feat(agent-memory): Phase 8 — self-optimization config reads#256
feat(agent-memory): Phase 8 — self-optimization config reads#256jamby77 wants to merge 4 commits into
Conversation
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
eea7699 to
ee647cc
Compare
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
524f3a8 to
e58dbb7
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
14b6e7c to
9412e5a
Compare
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
e58dbb7 to
0d89d67
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
9412e5a to
37c7a4b
Compare
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 37c7a4b. Configure here.
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
0d89d67 to
73d91d7
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
37c7a4b to
06b56f6
Compare
KIvanow
left a comment
KIvanow
left a comment
There was a problem hiding this comment.
Phase 8 is in good shape: off-by-default keeps existing stores quiet, the before-await snapshots are genuinely tear-free since applyConfig is synchronous and reassigns rather than mutates, the all-zero weights rejection is the right guard, and the config key stays clear of the index prefix. One thing I would like addressed:
A partial weights write silently resets the omitted components to constructor defaults. Because applyConfig rebuilds the weights vector from {...initialWeights} on every pass, a config hash that sets only recall.weights.similarity snaps recency and importance back to the constructor values (0.25 / 0.15) rather than preserving the currently-live ones. That is consistent with the documented snapshot-fallback semantics, but for Monitor's proposal engine, which is likely to nudge a single knob, it is a footgun: tuning one weight silently retunes the other two. Please either preserve the live weight components when only a subset is present, or document clearly that weights must always be written as a full triple so a partial proposal does not quietly reset the rest.
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
73d91d7 to
8bc0c13
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
7d0fa87 to
90ac526
Compare
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
8bc0c13 to
9f8485a
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
2 times, most recently
from
f325dee to
94e4101
Compare
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
94d2a7d to
66c0809
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
94e4101 to
ff2e1f1
Compare
- MemoryStore reads {name}:__mem_config on an opt-in configRefresh
interval and live-applies recall.threshold, recall.weights.*,
recall.halfLifeSeconds, and maxItemsPerScope without a restart
- Absent fields fall back to constructor values; invalid values are
ignored; reads are best-effort and never throw
- Reject an all-zero weight vector to keep recall ordering well-defined
- Add currentConfig() to expose the effective tunables; close() also
stops the refresh interval
- Refresh is off by default (opt-in) so a standalone store never polls
Capture halfLifeSeconds with threshold/weights before the first await so a concurrent configRefresh can't score a single recall with a mix of config versions.
enforceCapacity read weights and halfLifeSeconds from instance fields at selectEvictions time, several awaits after the capacity check. With opt-in configRefresh a refresh could land mid-pass and score victims with a different tunable set. Snapshot weights and halfLifeSeconds at entry, matching the recall snapshot, so a pass uses one consistent config.
…g write applyConfig rebuilt the weights vector from the constructor defaults each pass, so a config that set only one weight component silently reset the other two — a footgun for the proposal engine, which nudges a single knob. A partial write now starts from the live weights and overlays only what's present; a fully-absent weights config still falls back to the constructor values.
jamby77
force-pushed
the
feature/agent-memory-phase7-discovery
branch
from
66c0809 to
c83f1e6
Compare
jamby77
force-pushed
the
feature/agent-memory-phase8-config
branch
from
ff2e1f1 to
0efd8ba
Compare
2 participants
Stacked on #255 (Phase 7 — discovery marker).
What
Phase 8 of
@betterdb/agent-memory: runtime-tunable recall/eviction knobs, so BetterDB Monitor's cache-proposals engine can retune a memory store without a restart (same pattern as semantic-cache'sconfigRefresh).MemoryStorereads{name}:__mem_configand live-applies:recall.threshold(cosine-distance ceiling, 0..2)recall.weights.similarity/recall.weights.recency/recall.weights.importancerecall.halfLifeSecondsmaxItemsPerScoperefreshConfig()(public, manual tick) reads the hash and applies it;currentConfig()exposes the effective tunables.Opt-in
configRefresh?: boolean | { enabled?, intervalMs? }enables an immediate read plus an unref'd interval (default 30s, min 1s);close()stops it.Design / review notes
.callsequences.initialWeightsis copied so the sharedDEFAULT_WEIGHTSconstant can't be aliased.recall()captures threshold+weights into locals before its firstawait, so an interval refresh can't tear an in-flight recall.Tests (
MemoryStore.config.test.ts, 12)defaults snapshot · threshold · weights · halfLifeSeconds+maxItemsPerScope · partial config leaves others default · field-removal reverts · invalid ignored · all-zero weights rejected · live threshold affects recall end-to-end · no polling when disabled · immediate+interval read & stop on close (fake timers) · best-effort on read failure.
77/77 package tests green ·
tscclean · prettier clean.Note
Medium Risk
Changes live recall scoring, threshold filtering, and eviction victim selection when refresh is enabled; defaults stay off and invalid config is ignored, but mis-tuned remote hashes could alter memory behavior at runtime.
Overview
Adds opt-in runtime tuning for
MemoryStoreby reading{name}:__mem_configfrom Valkey, mirroring semantic-cache’sconfigRefreshpattern so Monitor can retune without a restart.refreshConfig()andcurrentConfig()expose manual reads and the effective snapshot.configRefresh(off by default) triggers an immediateHGETALLplus an unref’d interval (default 30s, min 1s);close()clears the timer. Tunables includerecall.threshold, partial recall weight fields,recall.halfLifeSeconds, andmaxItemsPerScope, with validation, best-effort reads, rejection of all-zero weights, and constructor fallbacks when fields are missing.recall()andenforceCapacity()snapshot half-life (and eviction weights) so a concurrent refresh cannot mix config versions mid-operation. NewMemoryStore.config.test.tsand an eviction test cover the behavior;index.tsexports the new config types.Reviewed by Cursor Bugbot for commit 0efd8ba. Bugbot is set up for automated code reviews on this repo. Configure here.