* fix: stop burning Claude Pro tokens on every tool call (#143)
0.8.8 fixed the agentmemory-side Claude API burn (where the engine
called Claude via the user's ANTHROPIC_API_KEY for per-observation
compression). That addressed #138 for users with API keys, but it
missed the second and much larger token-burn path: the PreToolUse
hook writing context to stdout.
Claude Code reads PreToolUse stdout and prepends it to the model's
next turn. src/hooks/pre-tool-use.ts was POSTing /agentmemory/enrich
on every Edit/Write/Read/Glob/Grep tool call and piping up to 4000
chars of response context into stdout. At ~20 tool calls per user
message this silently injected ~20K tokens per message into Claude
Code's input window — all charged against the user's Claude Pro
allocation because Claude Code was the one sending them to Anthropic.
4 messages drained the cap, which matches @adrianricardo's report.
session-start.ts had the same pattern (injected once per session,
smaller blast radius).
Fix: gate both hooks on AGENTMEMORY_INJECT_CONTEXT, default false.
- pre-tool-use.ts: when disabled, exit immediately — no stdin read,
no fetch, no stdout write. The hot path (~20x per message) becomes
a no-op Node startup.
- session-start.ts: when disabled, still POST /agentmemory/session/start
so the session gets registered for observation tracking, but never
write context to stdout. The session registration is cheap and
doesn't touch Claude Code's input window.
- src/config.ts: new isContextInjectionEnabled() helper.
- src/index.ts: startup banner prints 'Context injection: OFF
(default, #143)' on normal startup and a loud WARNING when
opt-in is enabled.
- test/context-injection.test.ts: 5 subprocess tests that spawn the
compiled pre-tool-use.mjs and session-start.mjs hooks, feed real
JSON payloads via stdin, and assert stdout is empty in all the
off/default paths. Also asserts the disabled path exits under 1s
and the opt-in path with an unreachable backend still exits
cleanly.
- README .env section: new AGENTMEMORY_INJECT_CONTEXT entry.
- CHANGELOG [0.8.10] with prominent 'Behavior change' banner.
Observations are still captured via PostToolUse regardless of the
flag — the memory store and MCP search tools are completely
unaffected by this change. The fix only severs the path where
agentmemory silently shoves memory context into the user's Claude
Code conversation.
Bumps to 0.8.10 (main + @agentmemory/mcp shim). Test count:
724 passing (was 719 + 5 new).
* chore: rewrite #143 CHANGELOG entry with corrected diagnosis
PreToolUse stdout is NOT injected into the model context — per
the Claude Code hook docs, only UserPromptSubmit and SessionStart
stdout are injected. My initial #143 PR description and CHANGELOG
claimed PreToolUse was the smoking gun behind 'Pro allocation
burned in 4 messages', which is wrong.
What's actually true:
- SessionStart stdout injection IS real (~1-2K tokens per session)
- PreToolUse stdout goes to debug log only — no tokens
- Claude Pro's Claude Code quotas are tight by design (Anthropic
has publicly acknowledged this); 4 messages to burn is plausible
with or without agentmemory installed
The gate on pre-tool-use.ts is still worth keeping as a resource
cleanup (skips a 20x-per-message Node+HTTP hot path) and as
forward-compat protection in case Claude Code ever changes
PreToolUse hook contract. But the CHANGELOG entry has to stop
claiming it saves tokens when it doesn't.
* fix: mem::retention-evict no longer leaks semantic memories (#124)
The eviction loop was unconditionally calling kv.delete(KV.memories,
id) for every below-threshold candidate, but retention scores are
computed for both episodic (KV.memories) and semantic (KV.semantic)
memories. When a candidate came from KV.semantic, the delete silently
became a no-op (key wasn't in mem:memories to begin with) and the
semantic row stayed alive forever with a sub-threshold score. Semantic
memories could not be evicted by this path at all.
Fix:
- Add a source: "episodic" | "semantic" discriminator to RetentionScore
- Tag it at score creation in both loops of mem::retention-score
- Branch the delete in mem::retention-evict on candidate.source,
routing to KV.memories or KV.semantic accordingly
- Pre-0.8.10 retention rows with no source field are treated as
episodic for backwards-compat so upgraded stores continue to evict
their old rows without re-scoring first
- Response now includes evictedEpisodic and evictedSemantic counts
so callers can see what was removed from each scope
Adds 3 regression tests to test/retention.test.ts:
- Scoring tags rows with the correct source
- Evicting a mixed set of below-threshold episodic + semantic
candidates removes both from their respective scopes
- Legacy-shape score rows with no source field still evict
to mem:memories (backwards-compat)
Full suite: 727 passing (was 724 + 3 new).
* review: probe namespaces for legacy retention rows (#124, round 2)
CodeRabbit caught a real backwards-compat hole in the #124 fix:
pre-0.8.10 stores already contain semantic retention rows with
no source field (because the old mem::retention-score scored
KV.semantic before the discriminator existed). My first fix
defaulted missing source to episodic, which meant those legacy
semantic rows still got delete-routed to KV.memories — the exact
no-op that stranded them in the first place.
Fix: when candidate.source is undefined, probe KV.memories first
for the memoryId; if it's there, route to episodic, otherwise
route to semantic. Count the resolved source in the response.
Adds one new test case: a pre-0.8.10 semantic memory with a
legacy-shape retention row (no source field) gets evicted from
mem:semantic, not silently no-op'd. Existing 'defaults to
episodic' test is kept and retargeted to the genuinely-episodic
legacy case.
Also fixes a README nit: the AGENTMEMORY_INJECT_CONTEXT comment
previously implied SessionStart fires on every tool turn. It's
once per session. Now broken out into two bullets explaining
what each hook does differently, with the note that only
SessionStart actually reaches the model (PreToolUse stdout is
debug-log only per Claude Code docs).
Full suite: 728 passing (was 727 + 1 new).
* review: audit retention evictions + assert persisted source (#124 round 3)
CodeRabbit round 3 findings, both real:
1. retention-evict performs structural deletes (memories / semantic /
retention scores / access logs) but was not calling recordAudit().
Repo learnings say state-changing functions must be auditable except
for read-path bookkeeping. Now emits one batched audit row per
non-zero eviction sweep:
operation: 'delete'
functionId: 'mem::retention-evict'
targetIds: every evicted memoryId
details: { threshold, evicted, evictedEpisodic, evictedSemantic,
reason: 'retention score below threshold' }
Zero-eviction sweeps intentionally do NOT write an audit row (no
state change, no need to flood the audit log during health checks).
2. The #124 scoring test only checked result.scores (transient
response) but not the persisted mem:retention rows. Eviction reads
back from stored rows, so a regression in kv.set/serialization
would have still passed the old assertion. Now also does
kv.get('mem:retention', id) and asserts { source: ... }.
Two new tests:
- Retention evict with a mixed set of 2 episodic + 1 semantic
candidates writes exactly one audit row with all 3 ids in targetIds
and the correct evictedEpisodic/Semantic breakdown in details.
- Retention evict with zero candidates writes zero audit rows.
Full suite: 730 passing (was 728 + 2 new).
* review: audit retention-score + parallelize writes (#124 round 4)
CodeRabbit round 3 outside-diff findings, both addressed:
1. mem::retention-score was persisting schema-relevant writes to
KV.retentionScores (1000+ rows in a mature store) but never called
recordAudit(). Per the repo audit-coverage policy, state-changing
functions need an audit row. Added a single batched audit event
per rescore:
operation: 'retention_score' (new audit op — added to the
AuditEntry union in types.ts)
functionId: 'mem::retention-score'
targetIds: [] (intentionally empty — a mature store can have
1000+ ids per sweep; flooding the audit log
with every memoryId on every cron tick is
worse than recording just the summary counts)
details: { total, episodic, semantic, tiers, config }
Zero-memory stores intentionally skip the audit call.
2. The per-memory kv.set inside the score loop was O(n) sequential
round-trips. Refactored to collect pendingWrites: [id, entry][]
while iterating, then flush with Promise.all at the end. On a
mature store with 1000+ memories this is ~10x faster (depends on
backend pipelining).
Test updates:
- Added 'mem::retention-score emits audit row per rescore' covering
the new audit call, targetIds=[], and details.episodic/semantic.
- Existing '#124 audit evict' and 'zero-evict skip audit' tests now
filter the audit log by functionId === 'mem::retention-evict'
because retention-score also writes one row per sweep now.
Full suite: 731 passing (was 730 + 1 new, existing tests retargeted).
* chore: add viral gist callout + crop banner whitespace
- README: new sub-line under the tagline crediting the design doc
gist at 669 stars / 91 forks, with a link. The gist extends
Karpathy's LLM Wiki pattern with confidence scoring, lifecycle,
knowledge graphs, and hybrid search — it's the spec agentmemory
implements, so readers get context on what they're looking at.
- assets/banner.png: cropped from 930x512 to 930x332 (removed
~70px dead space from the top and ~110px from the bottom).
The wordmark, tagline, orange underline, and top of the neural
network graphic are all intact. Aspect ratio is now ~2.8:1
which renders better in the GitHub repo header.
* readme: bold viral-gist callout — shields.io badge in banner-orange
Previous sub-line was buried as gray <sub>, practically invisible
next to the npm/tests/license badges. Replace with a for-the-badge
style shield in #FF6B35 (matches the orange underline in the banner)
so it catches the eye the moment the page loads. Caption stays as
<sub> underneath for people who want the context.
* fix: Claude Code plugin auto-wires MCP + sandbox-safe skills (#139)
Two fixes reported in #139:
1. The Claude Code plugin (rohitg00/agentmemory) only shipped hooks
and skills; it did not auto-register any MCP server. Users had to
configure the @agentmemory/mcp stdio server separately, which the
README didn't make clear. Add plugin/.mcp.json so
/plugin install agentmemory@agentmemory wires up the MCP server
automatically when the plugin is enabled.
2. The `recall` and `session-history` skills used inline bash
pre-execution blocks (`!$(curl -s ... ${VAR:-default})`) with
shell expansion, which Claude Code's sandbox rejects by pattern
match with "Contains expansion" — no way to bypass except for
Claude to reconstruct the curl call manually, as the reporter
had to work around.
Rewrite all four plugin skills (recall / remember / forget /
session-history) as pure prompts that tell Claude to use the MCP
tools directly: memory_smart_search, memory_save, memory_sessions,
memory_governance_delete. No bash, no sandbox issues, no shell
escaping. Skills also run faster — no curl subprocess fork on
every invocation.
Also tighten the README Claude Code install snippet to note that
the plugin now registers hooks + skills AND auto-wires the MCP
server in one step.
Bumps to 0.8.9. 707 tests still pass (no behavioral changes to
server code; only plugin assets).
* review: implement the MCP tools the new skills call (#139)
CodeRabbit flagged that the rewritten plugin skills reference
memory_smart_search and memory_governance_delete, but the
standalone MCP shim (@agentmemory/mcp) only exposed 5 tools:
memory_save, memory_recall, memory_sessions, memory_export,
memory_audit. At runtime the skills would have thrown
"Unknown tool" the first time they ran.
Add the missing handlers to src/mcp/standalone.ts:
- memory_smart_search — aliases memory_recall, falling back to
substring search over the in-memory KV store. The standalone
shim has no engine behind it, so BM25/vector/graph aren't
available, but the substring match is good enough for plugin
usage on small personal memory stores.
- memory_governance_delete — deletes memories by id. Accepts
memoryIds as an array or a comma-separated string, returns
{deleted, requested, reason}. Throws on missing/empty input,
silently skips unknown ids.
Also normalize memory_save concepts/files: they now accept both
arrays (what the new skills send) and CSV strings (what the old
skills and legacy clients send). New normalizeList() helper
centralizes the logic.
Fix memory_sessions to honour a limit arg (default 20) — it was
previously returning every session in the store.
Update the forget skill wording per CodeRabbit's nit: make it
explicit that memory_governance_delete only takes memoryIds, and
that deleting a whole session means collecting all memory IDs
from that session via search first.
Adds 8 regression tests to test/mcp-standalone.test.ts covering:
- memory_save with array concepts/files
- memory_save with CSV concepts/files (legacy path)
- memory_smart_search substring fallback
- memory_sessions limit
- memory_governance_delete array path
- memory_governance_delete CSV path
- memory_governance_delete missing/empty validation
- memory_governance_delete unknown-id skip
Full suite: 715 passing (was 707 + 8 new).
* review: reject empty search query + fix skill troubleshooting paths (#139)
CodeRabbit round 2:
1. src/mcp/standalone.ts memory_smart_search accepted '' as a
query and used .includes('') which matches every memory. In
the forget flow this would surface unrelated memory IDs for
deletion. Now throws 'query is required' on missing/empty/
whitespace-only input. Also broadens the search corpus from
just title+content to include files, concepts, sessionIds,
and id — so users can /forget by file path, concept name, or
session ID, as the skill docs already promise.
2. All four skill files (recall, remember, forget, session-history)
said 'check the agentmemory engine is running on localhost:3111'
when tools are missing. Wrong troubleshooting path: the stdio
MCP shim spawned by Claude Code via .mcp.json has no REST API,
no port, no localhost. The correct checks are /plugin list,
/mcp, and restarting Claude Code after install. Updated all
four skills.
Adds 2 new regression tests to test/mcp-standalone.test.ts:
- memory_smart_search rejects empty/whitespace/missing query
- memory_smart_search finds memories by file, concept, or id
Full suite: 717 passing.
* review: tighten content/limit validation in standalone MCP (#139)
CodeRabbit round 3:
1. memory_save cast args.content as string and called .trim(),
which throws a runtime TypeError if content is a number, object,
null, etc. Replace the cast with a proper typeof guard that
returns the clean 'content is required' error for any non-string
input.
2. memory_smart_search trimmed rawQuery for emptiness but then
lowercased the *untrimmed* value, so a query like ' bcrypt '
would fail substring matches with trailing whitespace. Trim
first, then lowercase.
3. (args.limit as number) || 10 is fragile — it treats 0 and NaN
the same, silently accepts negative/Infinity values, and has no
upper bound so a malicious or buggy caller could request a
million memories in one slice(). Add parseLimit() helper that:
- rejects non-number/non-string types (booleans, objects, arrays)
- rejects NaN, Infinity, and non-positive values
- clamps to MAX_LIMIT=100 via Math.min
- Math.floor()s to force integer
- falls back to a per-call default (10 for search, 20 for
sessions, 50 for audit)
Applied to all three limit sites: memory_smart_search,
memory_sessions, memory_audit.
Adds 2 regression tests:
- memory_save rejects 42 / {} / [] / null / undefined / true without
crashing on .trim()
- parseLimit falls back on bogus values and clamps 99999 -> 100
Full suite: 719 passing (was 717 + 2 new).
* fix: stop burning Claude tokens on every PostToolUse hook (#138)
mem::observe was firing mem::compress unconditionally on every
PostToolUse hook, which called Claude via the user's ANTHROPIC_API_KEY
to compress each raw observation into a structured summary. On an
active coding session (50-200 tool calls/hour) this silently burned
hundreds of thousands of Claude API tokens within minutes. Reported
by @olcor1 on a fresh Docker install — his token allocation was gone
in 20 minutes.
Make per-observation LLM compression opt-in.
- New env var AGENTMEMORY_AUTO_COMPRESS, default `false`. When false,
observe.ts builds a zero-LLM synthetic CompressedObservation from
raw tool I/O (title from toolName, narrative from truncated
input/output, files extracted from file_path/pattern/etc., type
inferred from camelCase-normalized tool name) and stores + indexes
it. Recall and BM25 search still work.
- New src/functions/compress-synthetic.ts with buildSyntheticCompression().
- src/functions/observe.ts gates the compress trigger on the flag and
writes the synthetic observation + stream events when disabled.
- Startup banner now logs "Auto-compress: OFF (default, #138)" or a
loud warning when opt-in is on, so the mode is never silent.
- Regression test: test/auto-compress.test.ts covers the default path
(no mem::compress fired), explicit opt-in, tool-name-to-type
mapping, file-path extraction, narrative truncation, and the
post_tool_failure→error mapping.
Bumps to 0.8.8 (main package + shim). 707 tests pass.
**Migration**: users who were relying on LLM-generated summaries set
AGENTMEMORY_AUTO_COMPRESS=true in ~/.agentmemory/.env and restart.
Existing compressed observations on disk are untouched.
* review: address CodeRabbit findings on #140
- compress-synthetic.ts:92 — use the fallback toolName for inferType
so both title and type derive from the same resolved value
- test/auto-compress.test.ts — vi.resetModules() in beforeEach to
prevent cached observe.js from seeing stale env var state between
the default, opt-in, and explicit-off test cases
docker-compose.yml mounts `./iii-config.docker.yaml:/app/config.yaml:ro`
but the `files` array in package.json only shipped `iii-config.yaml`.
Docker resolves missing bind sources by silently creating them as empty
directories, so `npx @agentmemory/agentmemory` mounted an empty dir at
/app/config.yaml and iii-engine crashed with:
Error: Failed to read config file '/app/config.yaml':
Is a directory (os error 21)
Fix: add iii-config.docker.yaml to package.json `files`.
Also adds a regression test in test/consistency.test.ts that parses
every `./<path>:<container>` bind mount in docker-compose.yml and
asserts the source file is shipped via the `files` array, so this
class of bug can't reach a release again.
Bumps main package + shim to 0.8.7 and adds CHANGELOG [0.8.7].
0.8.5's publish workflow got blocked on the unscoped `agentmemory-mcp`
name by npm's name-similarity policy (there's an unrelated third-party
package called `agent-memory-mcp`). Publishing the shim under the
`@agentmemory` scope we already own sidesteps the conflict.
Changes:
- Rename packages/agentmemory-mcp → packages/mcp
- Package name: agentmemory-mcp → @agentmemory/mcp (0.8.4 → 0.8.6)
- Keep bin name `agentmemory-mcp` so `npm i -g @agentmemory/mcp`
still installs a binary at `agentmemory-mcp` for muscle memory
- Log prefix [agentmemory-mcp] → [@agentmemory/mcp] in standalone.ts
and in-memory-kv.ts
- README / integrations/openclaw / integrations/hermes / src/cli.ts
all point at `npx -y @agentmemory/mcp`
- .github/workflows/publish.yml working-directory and npm view queries
target packages/mcp and @agentmemory/mcp
- Shim dependency on @agentmemory/agentmemory bumped to ~0.8.6
- Main package 0.8.5 → 0.8.6 across the 8 version files
- CHANGELOG [0.8.6] entry documenting the scope move
CodeRabbit flagged that processLine only checked `id != null` in the
invalid-shape path, which meant (a) malformed messages with non-primitive
ids could be echoed back with that id, and (b) valid-shape requests with
id: {} / [] / true would fall through to the handler and be responded to
with the bogus id as the response id — both spec violations.
Add isValidId() (string | number | null | undefined), extract rawId once
as unknown, and:
- In the invalid-shape path, only echo if rawId is string or number.
- After shape validation, reject bad-id requests with -32600 / id: null
before invoking the handler.
Adds 5 test cases covering id: object, array, boolean, plus the malformed
+ non-primitive-id drop path and a positive string-id case.
Codex CLI v0.120.0 fails to start the standalone MCP server with
"MCP startup failed: Transport closed". Other MCP clients (Claude
Desktop, Cursor, OpenClaw) work fine with the same binary. Reported
in #129 by @GuillaumeOuint.
Root cause: src/mcp/transport.ts unconditionally writes a JSON-RPC
response for every incoming message, including notifications. Per
JSON-RPC 2.0 §4.1 a notification is "a Request object without an id
member" and "the Server MUST NOT reply to a Notification". Strict
clients close the transport when they receive an unsolicited
response — exactly the symptom Guillaume reported.
The trigger sequence on Codex CLI:
1. Codex sends initialize (request, has id) → server responds ✓
2. Codex sends notifications/initialized (notification, no id)
3. Our handler in standalone.ts returns {}
4. Transport writes {"jsonrpc":"2.0","result":{}} to stdout
(no id field because request.id was undefined)
5. Codex parses an unsolicited response → closes transport
6. Guillaume sees "Transport closed"
This worked accidentally for lenient clients (Claude Desktop,
Cursor, etc.) which silently ignored the spurious responses.
Fix:
- transport.ts: detect notifications via missing/null id field,
run the handler for side effects, then drop the response. Errors
inside notification handlers are logged to stderr (not silently
swallowed) so production debugging still works
- transport.ts: malformed messages that have no id are also dropped
silently (consistent with notification semantics) instead of being
responded to with id: null
- transport.ts: extracted processLine() so the line-handling logic
is testable independently of process.stdin/stdout
- new test/mcp-transport.test.ts with 9 tests covering: request
responses (success + error), notification suppression (no id +
null id + handler-throws), malformed input (parse error, invalid
request shape with id, invalid request shape without id, empty
lines)
Tests: 693/693 (684 baseline + 9 new). Build clean.
CodeRabbit flagged two remaining minor issues on the Windows diagnostics
PR #128 that landed before they were addressed:
1. src/cli.ts:161 — the exit handler only treated non-zero numeric
codes as abnormal. Signal-killed processes have code === null and
signal !== null (e.g., SIGSEGV, SIGABRT, SIGKILL). Those crashes
were silently dropped from startupFailure so the user got the
generic "did not become ready" timeout instead of a real diagnosis.
On Windows this rarely fires (abnormal exits usually carry codes)
but on Unix a segfault in iii-engine would be invisible.
Fix: widen the guard to `(code !== null && code !== 0) || (code ===
null && signal !== null)` so signal-only exits are captured too.
Also generate a clearer fallback message when stderr is empty:
"process killed by signal SIGSEGV" rather than the old "process
exited with code null (SIGSEGV)".
2. src/cli.ts:114 — when USERPROFILE (Windows) or HOME (Unix) is
unset, `join("", ".local", "bin", "iii.exe")` produces a relative
path like `.local/bin/iii.exe`. Harmless today (the path is checked
via existsSync which fails silently), but theoretically it could
match an unintended file in the user's current working directory.
Fix: guard the env var first, return an empty list (Windows) or
skip the per-user path (Unix) when the env var is missing. Keep
the absolute /usr/local/bin/iii fallback on Unix since it's valid
regardless of HOME.
No test changes — whichBinary and fallbackIiiPaths are CLI side-effect
paths not covered by unit tests. Tests: 684/684. Build clean.
CodeRabbit caught: `execFileSync(...).trim().split("\n")[0]` leaves a
trailing \r on the first path whenever `where` on Windows returns
multiple matches. `.trim()` only strips leading/trailing whitespace —
the CRLFs between the matches stay, so split("\n")[0] yields
"C:\path\one.exe\r" which spawn() rejects as an invalid path.
Single-match outputs worked by accident because .trim() ate the final
\r\n before the split. The bug only fired when a user had iii installed
in two places simultaneously (e.g. a WSL install alongside a native
Windows install), which is a plausible but non-default setup.
Fix: split on /\r?\n/, trim each line, find the first non-empty one.
Also now returns `null` on empty output instead of the empty string
(behavior is identical through the caller's truthy check, but it
matches the documented Promise<string | null> contract exactly).
Caught during PR #128 review: the install hints I added showed
`cargo install iii-engine` as the recommended path on both Windows
and Unix. That crate does not exist on crates.io. Verified via:
- https://crates.io/api/v1/crates?q=iii (no iii-engine result)
- https://crates.io/api/v1/crates/iii-engine (404)
The actual install methods per github.com/iii-hq/iii releases:
- macOS / Linux: curl -fsSL https://install.iii.dev/iii/main/install.sh | sh
(downloads the prebuilt binary into ~/.local/bin/iii)
- Windows: download iii-x86_64-pc-windows-msvc.zip from the GitHub
release page, extract iii.exe, add to PATH or place at
%USERPROFILE%\.local\bin\iii.exe (which our fallback already checks)
- Docker: docker pull iiidev/iii:latest (already used by docker-compose)
There is no PowerShell installer (https://install.iii.dev/iii/main/install.ps1
returns 404) and no scoop / winget / cargo path on Windows. The previous
README "From source" line (`cargo install iii-engine`) was already wrong
on main — fixed at the same time so we have one correct source of truth.
Changes:
- src/cli.ts installInstructions(): rewrite both Windows and Unix
branches to show the correct upstream commands. Add link to
https://iii.dev/docs at the bottom of each
- src/cli.ts troubleshooting note: "iii-engine version mismatch" line
no longer references cargo
- src/cli.ts fallbackIiiPaths(): drop the .cargo/bin/iii.exe fallbacks
on both platforms — they were dead paths since the crate doesn't
exist. Keep .local/bin/iii.exe (matches where the sh installer puts
it on Unix and where the README tells Windows users to drop iii.exe)
- README.md "Windows" section: rewrite Option A as "download the
prebuilt MSVC zip from GitHub releases", add explicit disclaimer
note that there is no cargo install iii-engine path
- README.md "From source" line: replace stale cargo command with the
correct sh-installer / Windows-zip / Docker matrix
Same test count (684 / 684), build clean.
Root cause: when the iii-engine binary crashes on startup or isn't
installed at all, the CLI spawned with `stdio: "ignore"` and swallowed
stderr entirely. 15 seconds later it reported "Check that ports 3111,
3112, 49134 are available" — which pointed users at the wrong problem.
On Windows this was the most common failure mode because most users
don't have Rust/cargo installed and nothing tells them they need it.
Changes:
- Capture engine stderr via `stdio: ["ignore", "ignore", "pipe"]` with
a 16KB cap, surface it on startup failure or in --verbose mode.
Processes still spawn detached/unref'd so the parent can exit cleanly.
- `windowsHide: true` so spawning a console app doesn't pop a window.
- Platform-aware iii binary fallback paths: on Windows, search
%USERPROFILE%\.cargo\bin\iii.exe, %LOCALAPPDATA%\cargo\bin\iii.exe,
and %USERPROFILE%\.local\bin\iii.exe. On Unix, add ~/.cargo/bin/iii
alongside the existing paths.
- Use `path.delimiter` instead of hardcoded `:` when prepending to PATH.
- Three-way diagnostic when the engine never becomes ready:
1. Process crashed → show the captured stderr + common causes
2. Process never started → platform-specific install instructions
3. Running but port not responding → platform-specific port check
command (netstat on Windows, lsof/ss on Unix)
- New `--verbose` / `-v` flag echoes spawn commands and surfaces engine
stderr even on success. Useful for diagnosing #127-style reports
across any platform.
- Docker compose lookup searches three candidate paths (was two), fixing
a latent bug where npm-installed docker-compose.yml at the package
root wasn't always found.
- README: new `### Windows` section in the install area documenting
Option A (rustup + cargo install iii-engine), Option B (Docker
Desktop), and Option C (standalone MCP only, no engine), plus a
symptoms→fix troubleshooting table.
No test changes needed — engine startup is a side-effect path exercised
only via the CLI binary itself, not imported by any unit test. Verified
via `npm test` (684/684) and `npm run build` (clean).
Two community contributions on top of 0.8.3 land as a patch release so
they reach users immediately and so `npx agentmemory-mcp` ships to the
npm registry for the first time (closing out #120 on the live release).
- #122 (Jason Landbridge): memory_save persists to disk immediately
after every call, not just on SIGINT/SIGTERM, so memories survive
SIGKILL from agent sessions ending forcefully
- #121 (Jason Landbridge): OpenCode MCP config format corrected —
opencode.json with mcp/type:local/command array instead of the
wrong .opencode/config.json with mcpServers
Version bumped across package.json, src/version.ts, types.ts,
export-import supportedVersions, plugin.json, shim package.json and
its dependency on the main package (~0.8.4). package-lock.json
regenerated. 684 tests passing, clean build.
* fix: persist memories to disk immediately after memory_save
Previously, kv.persist() was only called on SIGINT/SIGTERM. If the
process was killed with SIGKILL, crashed, or the host terminated it
forcefully (e.g. OpenCode session end), all saved memories were lost.
Fix: call kv.persist() immediately after kv.set() in the memory_save
handler so data is written to standalone.json on every save.
Also export handleToolCall with an injectable kv parameter to allow
unit testing, and add tests verifying that persist is called on save.
* refactor: use generateId() and single isoNow in memory_save handler
Replace ad-hoc ID/timestamp logic with project-standard helpers:
- generateId("mem") instead of manual Date.now()/Math.random() concatenation
- Single const isoNow = new Date().toISOString() shared by createdAt and updatedAt
fingerprintId is intentionally not used: memory_save always creates a new
unique entry (no dedup), so a content-addressable key would silently overwrite
duplicate content rather than append — matching existing project behaviour.
* fix: retention scoring (#119) + npx agentmemory-mcp 404 (#120), release 0.8.3
Issue #119 — retention score ignored agent-side reads
`mem::retention-score` previously hardcoded `accessCount=0` and
`accessTimestamps=[]` for episodic memories and used only a single-sample
`lastAccessedAt` for semantic memories, so the time-frequency decay
formula was a dead path. All read endpoints now record access events to
a new `mem:access` KV namespace via a keyed-mutex'd access tracker:
- mem::search, mem::smart-search (compact + expand), mem::context,
mem::timeline, mem::file-context, mem::get-related, mem::working-context
- retention.ts pre-fetches the full access log in one kv.list call,
normalizes every row defensively, and builds an O(1) lookup Map
- recordAccessBatch uses Promise.allSettled so a slow keyed-lock on one
id does not block the rest of the batch
- Bounded ring buffer of the last 20 access timestamps per memory
- Backwards-compat fallback: pre-0.8.3 semantic memories with only the
legacy sem.lastAccessedAt still score correctly
- Corrupt lastAccessedAt values are NaN-safe via Date.parse + isFinite
- reinforcementBoost is reported as the raw formula output, not the
clamped residual
Memory deletion sites (retention-evict, evict, auto-forget,
governance-delete, governance-bulk, forget, import replace) now call
deleteAccessLog to prevent mem:access from accumulating zombie entries.
mem::export / mem::import and mem::snapshot-create / mem::snapshot-restore
round-trip the new namespace so backup/restore cycles no longer zero out
reinforcement signals.
Issue #120 — npx agentmemory-mcp returned npm registry 404
The README instructed users to run `npx agentmemory-mcp`, but
`agentmemory-mcp` only existed as a `bin` entry inside
`@agentmemory/agentmemory`, not a real registry package. Two-part fix:
- New sibling package `packages/agentmemory-mcp/` — a thin shim that
depends on `@agentmemory/agentmemory` (~0.8.3, tilde to block 0.9.x
supply-chain drift) and forwards to dist/standalone.mjs via the new
`exports` field on the main package
- Canonical `npx @agentmemory/agentmemory mcp` subcommand on the CLI
for users who already have the main package installed
- Removed the dead `agentmemory-mcp` bin from the main package so the
shim owns the name unambiguously
- publish.yml publishes both packages in order with idempotent
guards and npm-view registry polling (no more fixed sleep)
- All MCP install snippets in README + integrations/openclaw + hermes
+ shim README use `npx -y agentmemory-mcp` to skip install prompts
Release 0.8.3
- 670 tests passing (665 + 11 new: access tracker, retention,
smart-search integration, corrupt-input, kv.list failure path,
Promise.allSettled resilience)
- Version bumped across package.json, src/version.ts, types.ts,
export-import supportedVersions, plugin.json
- package-lock.json regenerated
* fix: address CodeRabbit review feedback on #123
Real issues caught and fixed:
- access-tracker: deleteAccessLog now takes the same mem:access:<id>
keyed mutex used by recordAccess, eliminating the race where a
concurrent delete could drop an in-flight RMW update and leave a
stale access log for an evicted memory
- access-tracker: normalizeAccessLog now coerces count to a
non-negative integer, drops non-finite recent[] entries, truncates
recent[] to the last 20 on read (enforcing the ring-buffer invariant),
and guarantees count >= recent.length so a malformed row can't produce
NaN downstream. count is NOT capped at 20 — it is the lifetime access
counter, while recent[] is the 20-entry window
- export-import: import path now rejects accessLogs arrays above
MAX_ACCESS_LOGS (50_000), normalizes each entry via
normalizeAccessLog, and drops entries whose memoryId is not in the
imported memories set, so a crafted export can't blow up memory or
resurrect zombie access rows for memories that don't exist
- publish.yml: added a post-publish npm view polling loop for
agentmemory-mcp mirroring the one that already guards the main
package, so the workflow only succeeds when `npx agentmemory-mcp` is
actually visible on the registry
- CHANGELOG.md: added the missing [0.8.3] reference-style link target
at the bottom to match the [0.8.2]/[0.8.1]/[0.8.0] pattern
- README.md: standalone MCP inline examples now use `npx -y` to match
the MCP config snippets and skip the install confirmation prompt
Tests: 680/680 (+10 new — deleteAccessLog behavior, normalizeAccessLog
invariants, malformed mem:access row handling in retention scoring).
Findings deliberately not applied:
- auto-forget.ts / remember.ts / evict.ts recordAudit calls: existing
pattern in the codebase only audits governance deletes (user-initiated
admin actions), not automatic eviction. Adding audit everywhere is a
policy change, out of scope for this PR
- evict.ts gating deleteAccessLog on memory delete success: current
order is defensible, deleteAccessLog is already best-effort with its
own try/catch, and calling it after a failed memory delete is harmless
- snapshot.ts fail-fast on kv.list failure: inconsistent with the
existing .catch(() => []) pattern used for observations in the same
function
- retention.ts semantic evict only deletes KV.memories rows: real bug
but pre-existing on main (RetentionScore has no source/type field).
Tracked separately — out of scope for this PR
- bin.mjs fallback command: already present in the existing catch
handler
- Promise.all parallelization of delete+deleteAccessLog: micro
optimization not worth the readability cost
OpenCode uses a different MCP config format than other tools:
- Config file is opencode.json (not .opencode/config.json)
- Uses "type": "local" with "command" array (not mcpServers format)
- Does not use separate "command"/"args" fields
Also added an OpenCode-specific config block in the Standalone MCP
section alongside the generic JSON example so it's easy to find.
npx agentmemory-mcp fails because there is no standalone npm package
named 'agentmemory-mcp'. The binary is registered inside the
@agentmemory/agentmemory package and must be invoked with:
npx --package @agentmemory/agentmemory agentmemory-mcp
Updated all documentation examples across README.md and integration
guides (openclaw, hermes) to use the correct invocation. This fixes
MCP clients that followed the docs and got an npm 404 error.
Adds a full set of light-background SVG tags in assets/tags/light/
that mirror the existing dark-background tags in assets/tags/.
GitHub <picture> element auto-swaps based on the reader's theme:
- GitHub LIGHT theme → dark-bg SVG (original, stands out on white)
- GitHub DARK theme → light-bg SVG (new, stands out on dark)
## 30 light variants generated
Same dimensions, same orange #FF6B35 brand accent, same fonts. Only
the backgrounds and text colors change:
- Background: white → light gray gradient (#FFFFFF → #F8F9FB)
- Border: #E5E7EB (GitHub-style neutral border)
- Title text: #0F172A (slate-900)
- Subtitle text: #64748B (slate-500)
Stat values are slightly darkened for better contrast on white:
- Green: #00D26A → #16A34A
- Orange: #FF6B35 → #EA580C
- Purple: #B5A4FF → #7C3AED
## README changes
All 26 tag references now use <picture> elements:
<picture>
<source media="(prefers-color-scheme: dark)" srcset="assets/tags/light/foo.svg">
<img src="assets/tags/foo.svg" alt="..." height="..." />
</picture>
Covers: 14 section headers (both stat rows × 6 = 12 stat cards).
No visual changes for light-theme readers — the original dark SVGs
still show. Dark-theme readers now get proper contrast instead of
dark-on-dark blend.
All 30 light SVGs validated with xmllint. 654/654 tests passing.
* feat: flashy README with real agent logos + paste-this-prompt blocks for integrations
## Main README — real logos in Supported Agents grid
Replaced plain text cells with real brand logos for all 16 supported agents:
- simple-icons CDN for: Claude Code (Anthropic), Cursor, Gemini CLI
(Google Gemini), Codex CLI (OpenAI), Windsurf, Claude Desktop, Claude SDK
- GitHub org avatars for: OpenClaw, OpenCode (opencode-ai), Cline,
Hermes (NousResearch), Goose (block), Kilo Code (kilo-org), Aider
(Aider-AI), Roo Code (RooCodeInc)
- Shield badge for the generic "Any agent" cell
Each logo is clickable and links to the upstream project. The OpenClaw and
Hermes cells now link directly to integrations/openclaw and integrations/hermes
so users can find the deeper plugin integration.
Added tagline footer below the grid reinforcing the cross-agent story.
## Integration READMEs — paste-this-prompt blocks
Following the gstack pattern, both integration READMEs now open with a
visual hero (upstream logo + agentmemory logo) and a copy-pasteable text
block the user can drop into their agent. The agent handles the entire
setup — start server, update MCP config, verify health, open viewer.
This matches the "paste this into Claude Code" block we already have in
the main README for Claude Code users.
### openclaw/README.md
- Centered hero: OpenClaw org avatar + agentmemory icon
- Shield badges: 43 MCP tools, 4 lifecycle hooks, 95.2% R@5, self-hosted
- "Install it in 30 seconds" section with the paste-this prompt
### hermes/README.md
- Centered hero: NousResearch avatar + agentmemory icon
- Shield badges: 43 MCP tools, 6 lifecycle hooks, 95.2% R@5, self-hosted
- "Install it in 30 seconds" section with the paste-this prompt
All 654 tests still passing.
* fix: reorder agents, replace broken logos, add OpenClaw/Hermes paste blocks, cleaner hero
## Main README
### Supported Agents grid — 3 fixes
1. **Reorder**: OpenClaw and Hermes now in positions 2 and 3 (right after
Claude Code), since they have first-class gateway plugins in
integrations/. Users shouldn't have to scroll to find them.
2. **Fix broken Codex logo**: simple-icons CDN returned 404 for the
OpenAI slug with black color. Switched ALL logos to GitHub org
avatars (github.com/<org>.png?size=120) which are bulletproof and
always return real images. Verified 7/7 resolve to actual PNGs.
3. **Bigger logos**: bumped from 40px to 48px for more visual weight.
### Paste-this-prompt toggles
Added two collapsible <details> blocks after the Claude Code paste
prompt: one for OpenClaw, one for Hermes. Each contains a complete
copy-pasteable setup prompt that the agent can execute, plus a link
to the full integration guide.
## Integration README heroes — redesigned
The old hero had two logos separated by an ugly gray "+" shield badge
that looked cheap. Replaced with:
- agentmemory banner.png as the centered hero image
- H1 title with a small inline upstream logo (28px) next to the title
- Bolded tagline + subtitle with upstream link
- Shield badge row (MCP tools, hooks, R@5, self-hosted, license)
Much cleaner and follows the main README's visual pattern.
Verified: 654/654 tests passing.
* Update README.md
* feat: custom SVG tags for section headers (no emojis)
Replaces emoji section headers with custom SVG badges matching the
agentmemory brand style (orange #FF6B35 accent on dark #1A1A1A).
Inspired by shanraisshan/claude-code-best-practice's custom tag
approach in its !/ directory.
## Custom SVG tags created (30 total in assets/tags/)
### 14 section headers (section-*.svg)
Each is a 44px-tall dark rounded badge with:
- Orange gradient accent strip on the left (brand color #FF6B35 → #FF8F5E)
- Uppercase title in white, 800 weight, 2.2 letter-spacing
- Light gray subtitle (#9CA3AF) with section context
- Subtle #2A2A2A border on dark #1A1A1A → #0F0F0F gradient
Sections: works-with-every-agent, benchmarks, vs-competitors, quick-start,
why, how-it-works, search, mcp-server, real-time-viewer, configuration,
api, architecture, development, license.
### 6 hero stat blocks (stat-*.svg)
Big numbers with small uppercase labels, replacing the shield.io badge
row in two places (hero + benchmarks section):
- stat-recall: 95.2% R@5 (green)
- stat-tokens: 92% fewer tokens (green)
- stat-tools: 43 MCP tools (orange)
- stat-hooks: 12 auto hooks (orange)
- stat-deps: 0 external DBs (purple)
- stat-tests: 654 tests passing (green)
### 8 pill tags (pill-*.svg)
Rounded pills with colored dot indicator for inline status tags:
MCP, AUTO HOOK, SKILL, NEW, BETA, STABLE, PLUGIN, SECURE.
### 1 "New in v0.8.2" version tag
Dark pill with orange v0.8.2 version marker.
### 1 section divider
Horizontal line with orange diamond in the middle (unused for now,
available for future sections).
## README changes
- All <h2> section headers now use <img> tags pointing to the SVG
- Hero stats row replaced with stat-*.svg row (same visual position)
- Benchmarks stats row replaced with the same stat-*.svg row
- Nav links updated to match the new h2 anchor IDs
- "New in v0.8.2" callout uses new-v082.svg tag + note
- Cursor grid cell fixed (was missing <strong>Cursor</strong> label)
- Zero emojis in the entire README
All 30 SVGs validated with xmllint, 654/654 tests passing.
* Remove version 0.8.2 image from README
Removed image tag for version 0.8.2 from README.
* feat: add demo command, benchmark comparison, OpenClaw integration, token cost in viewer
Product & ecosystem improvements surfaced from the last30days research
and competitive analysis of the AI agent memory space.
## New
### agentmemory demo command
Seeds 3 realistic sessions (JWT auth, N+1 query fix, rate limiting)
and runs smart-search queries against them. Shows token-aware semantic
search finding "N+1 query fix" when you search "database performance
optimization" — proof that keyword matching can't do that.
npx @agentmemory/agentmemory demo
Surfaces the "show don't tell" moment in 30 seconds without requiring
the user to integrate anything.
### benchmark/COMPARISON.md
Head-to-head comparison vs mem0 (53K stars, 68.5% LoCoMo),
Letta/MemGPT (22K stars, 83.2% LoCoMo), Khoj (34K stars),
claude-mem (46K stars), and Hippo. Feature matrix across 18
dimensions. Token efficiency table. Honest caveats about LongMemEval
vs LoCoMo apples-vs-oranges. Lets users compare without trusting our
README.
### integrations/openclaw/
OpenClaw gateway plugin following the same pattern as the Hermes
integration. README, plugin.yaml with hooks config, and plugin.mjs
implementing the 4 lifecycle hooks (on_session_start, on_pre_llm_call,
on_post_tool_use, on_session_end). Gets agentmemory into the OpenClaw
ecosystem for distribution.
## Improved
### Viewer token savings card
Now shows tokens saved AND dollar cost saved (using $0.30/1K as a
reasonable mid-tier model rate) instead of just raw numbers. Makes
the ROI immediately tangible.
### README additions
- "Try it in 30 seconds" section featuring the new demo command
- Competitor comparison link in benchmarks section
- OpenClaw integration link in supported agents table
* refactor: simplify demo command and OpenClaw plugin
- Extract postJson helper in cli.ts runDemo (4x repeated fetch boilerplate → single call site)
- Break runDemo into buildDemoSessions, seedDemoSession, runDemoSearch for single-responsibility
- Replace three-branch if/else command dispatch with a commands map (single error handler)
- Extract typed DemoSession/DemoObservation/SearchResult interfaces
- Simplify OpenClaw plugin.mjs: collapse headers() + fetchJson() into a single postJson method
Each hook is now ~5 lines instead of ~10
No behavior change. 654/654 tests passing, CLI help and error paths verified.
* fix: address CodeRabbit review on growth changes
## plugin.mjs
- Honor the enabled config flag: constructor now sets this.enabled and
all four hooks (onSessionStart, onPreLlmCall, onPostToolUse,
onSessionEnd) return early when disabled
- Fail-fast on non-OK HTTP: previously !res.ok returned null regardless
of fallbackOnError. Now throws with status/statusText/body when
fallbackOnError is false
- Update header JSDoc to use the real hook names instead of the
fictional prefetch/capture/consolidate
## openclaw README.md
- Replace stale hook names (prefetch/capture/consolidate) with the
actual onSessionStart/onPreLlmCall/onPostToolUse/onSessionEnd names
matching plugin.mjs
- Add 'text' language identifier to the Project profile code fence to
satisfy markdownlint MD040
## cli.ts
- Replace ad-hoc Date.now()-based session IDs with generateId("demo")
from src/state/schema.ts, matching the rest of the codebase
(snapshot.ts, migrate.ts, consolidation-pipeline.ts)
- Add postJsonStrict helper that throws on non-OK with status + body
- seedDemoSession now uses postJsonStrict for session/start and
session/end (critical lifecycle calls that should fail fast)
- The observation POST in the loop now inlines the fetch to surface
HTTP errors via p.log.warn with status/body, still only incrementing
the stored counter on success
## viewer/index.html
- Fix token-to-dollar conversion that was 100x underreporting:
tokensSaved / 1000 * 0.3 returns DOLLARS, not cents. Now computes
dollars first, then cents = Math.round(dollars * 100)
- 100K tokens saved now correctly displays $30.00, not 30ct
All 654 tests still passing.
* fix: address 2 more CodeRabbit findings on plugin.mjs
- Use nullish coalescing (??) for config defaults so valid falsy
values like min_confidence: 0 or timeout_ms: 0 aren't overwritten
by defaults. || would clobber them.
- Fix /agentmemory/context payload to match the server contract at
src/triggers/api.ts:115. The handler expects { sessionId, project,
budget? } but the plugin was sending { sessionId, query, tokenBudget,
minConfidence }. Three issues:
1. Missing required 'project' field
2. Wrong field name: 'tokenBudget' should be 'budget'
3. 'query' and 'minConfidence' are not part of this endpoint's
contract (probably belong on /smart-search instead)
Fixes 6 security issues in agentmemory's default deployment, viewer, mesh sync, export path, and startup flow.
## Findings Addressed
1. **Default REST/stream exposure on 0.0.0.0** — bind default REST and stream services to 127.0.0.1; restrict Docker port publishing to localhost; add dedicated iii-config.docker.yaml for Docker usage.
2. **Stored XSS via inline viewer handlers** — remove inline event handlers from the viewer; replace with delegated data-action event handling; switch to nonce-based CSP for viewer HTML; add script-src-attr 'none'.
3. **Arbitrary path writes in Obsidian export** — confine exports to AGENTMEMORY_EXPORT_ROOT (default ~/.agentmemory); reject vaultDir values outside the allowed root.
4. **Mesh sync without peer auth** — require AGENTMEMORY_SECRET for mesh endpoints and sync operations; send Authorization: Bearer <secret> on mesh push/pull requests.
5. **Secret redaction misses modern token formats** — add coverage for Bearer tokens, sk-proj-*, broader sk/pk/rk/ak variants, GitHub ghs_/ghu_ tokens.
6. **CLI executes remote shell script directly** — remove curl|sh installation pattern; use existing local iii-engine binary if available, otherwise fall back to Docker.
## Validation
- 654 tests passing (8 new security tests)
- npm run build succeeds
- Merged with main and resolved README conflicts to preserve the new benchmark/competitor table layout
Thanks @eng-pf for catching these and submitting the fixes.
Build script now copies src/viewer/index.html to dist/viewer/.
Also added more candidate paths for index.html resolution so
the viewer works from dist/ (npx), src/ (tsx dev), and repo root.
Replace numbered steps with a single copy-pasteable instruction block
for Claude Code. The agent reads it and handles: engine startup, plugin
install, hook registration, and verification.
Other agents consolidated into a clean table (Cursor, Gemini CLI,
OpenCode, Claude Desktop, SkillKit, REST API) with inline config.
Rohit Ghumare