Skip to content

@staix/agent-mem v0.25.0

Choose a tag to compare

@github-actions github-actions released this 12 Nov 11:47
· 30 commits to main since this release
a981a2d
  • Add the person-owned preference note kind (amem note add --kind preference, amem_remember with kind: "preference"): a standing instruction that is long-term, project or shared scope, superseded and forgotten like any note, and never inferred or merged by amem consolidate. Briefs show it by its own slot instead of path- or intent-ranked rows: amem_brief reserves at most half its budget for up to five current preferences (this project's plus every shared one, newest first) and emits whole lines of at most 240 bytes, dropping one that does not fit and reporting truncated rather than clipping a rule in half; the SessionStart hook carries the same slot from its own prebuilt sidecar brief-preferences.json, written beside brief-curated.json under the same profile binding, 24-hour freshness and 32 KiB cap, at most 480 bytes, before the curated rows. A preference never takes a curated row, so one is never shown twice. Refs #370.

  • Compatibility with N-1 (0.24.0), measured: a preference row is not additive for an older reader. The published 0.24.0 reading a home whose own segment holds one lists 0 rows and returns syncMemory { ok: false, pushed: false, entries: 0 } with "this machine's long-term segment is not readable as complete valid rows, so it was not screened; inspect the local input and reopen the store before retrying". So the write fails closed: note add, importNote and accepting a preference candidate are refused until the memory home records {"surface":"curated","minReader":"0.25.0","since":"<date>"} in formats.json, and at that floor 0.24.0 refuses the home at store open with code: "format_floor" instead of breaking its sync, as measured on the same home. Reads, briefs and amem_remember queueing need no floor; a queued candidate cannot be accepted until the floor is raised, and the refusal leaves it pending. amem never writes formats.json.

  • Compatibility with N-1 (0.24.0), indexes and brief cache: INDEX_VERSION 4 names its own curated index file (index-v4.db, and index-v2-i4.db for the v2 public-prune view), so each release keeps its own warm index on one home and neither drops the other's; an older index skips preference rows, so sharing one file would be wrong. The cost is one refresh for a reader whose own file this release no longer advances: 0.24.0 rebuilds from the log and answers, and a reader that predates the vault generation meets the same refusal it already had for an op it does not know. The preference sidecar uses a new file name and metadata key, so 0.24.0 neither parses nor invalidates it, and the curated section it does read now excludes preference rows. Existing rows, hashes, schemas and paths are unchanged.

  • Extract canonical capture-only Records from positively completed main-session Codex, Pi and Kimi messages through the existing staging, receipt and replay path. Respect records: false, existing forward cursors, privacy, anchor and brief caps; add numeric JSON counts. Qwen and Jcode stage no Records without a native finality witness and report finality-unavailable; their ordinary capture remains unchanged. Refs #369.

  • Compatibility with N-1 (0.24.0), capture spool: non-Claude Records use the additive format-4 hook-v4- prefix, which 0.24.0 leaves pending while draining unchanged format-2/3 rows. The new reader drains all three formats with the same receipts and dedupe. Existing Claude publication bytes, curated schemas and floors are unchanged.

  • Select the hosted macOS StopBurst RSS regression baseline by CPU hardware class: keep 93.58 MB for M1 and unknown CPUs, and use the qualified M2 Pro cohort median of 103.19 MB. Retain the strict +10% gate and 150 MB ceiling; receipts name the class and applied baseline. Other groups and runners are unchanged. Refs #365.

  • Propose curated candidates from capture Record rows in amem consolidate: the newest 32 main-session Records of the project, grouped by kind and anchor set (by kind and text when the Record carries - anchors), one candidate citing up to 8 rows, with decision/contract/finding keeping their own note kind, fix becoming a finding titled fix: and fail filling symptom and retry-when from the two recorded groups. Additive only: the candidate input gains optional about (the anchors as path: and symbol: edges) and optional record (group key, anchors, and each cited row's session hash, time and line hash), the report gains a records block (found, queued, grouped, alreadyQueued, alreadyReviewed, excluded, refused, truncated), and acceptance point-reads the cited turns, so a candidate that outlives the 32-row window still proves its rows instead of reading them as pruned. canonical() drops undefined, so every existing candidate id, payload and hash is unchanged, and no index version, capture format or format floor moves; a host that passes no records callback consolidates exactly as before. Refs #368.

  • Compatibility with N-1 (0.24.0), candidates: the candidate store schema and every existing candidate id are unchanged. By code reading, 0.24.0 has no cited-row check and builds an accepted note's edges from its target alone, so accept Record-derived candidates with 0.25.0, which proves the cited rows and keeps the Record's path: and symbol: edges.

  • Compatibility with N-1 (0.24.0), root exports: the root now re-exports src/memory/record-candidates.ts (Record proposal constants, types and helpers); this is additive and existing export names remain. Hosts must pin 0.25.0 before calling them.