Skip to content

v0.13.0

Latest

Choose a tag to compare

@github-actions github-actions released this 18 Nov 16:19

v0.13.0 — complete evidence across similar project histories

This version improves the selection of original current-policy and boundary
evidence in long engineering histories. It preserves exact project/source
coordinates, handles a bounded set of bilingual declaration forms, and reduces
repeated directory work. The measured tradeoff is explicit: evidence coverage
improves, directory navigation is faster in every measured group, and many cold
recall groups are slower.

What changed

  • Qualify repeated declarations only when every literal mention independently
    establishes the same field, scope and value. Quoted, conditional, negated or
    unrelated text cannot silently establish a current decision.
  • Recognize bounded Chinese declaration and English reporting forms while
    preserving actual field names, namespace boundaries, occurrence queries,
    environment qualifiers and revocation controls.
  • Keep current declarations above authenticated no-change reports only when
    doing so preserves the requested facts. A query about a checklist result or
    review action retains that report's own source, including later changes.
  • Return bounded evidence assessment and conflict pointers, including compact
    MCP guidance. These describe the reached candidate pool; they do not certify
    whole-archive completeness, factual authority or safe abstention.
  • Batch directory search and current-revision candidate work, bound useful
    project row ranges, and use a guarded single-cursor membership filter for
    small interleaved projects. Original global FTS ranking remains in use;
    chapter/session paths retain their established planner behavior.
  • Prepare query/declaration terms after a validated result-cache miss. Project
    assertions, revision checks and other correctness checks remain in force.

Exact source and reproducible experiment

The source decision
accepts runtime SHA-256
a7492a74a94170b87a2fd54ada421866b9d737f46dce2ff3a3bc21d6debb9da1,
measured at e0e2515f6e16404cfb9b676cd2eab01ce84b81cc. Release preparation
retains those runtime bytes. The baseline is
8044e4bafe18e73ea422af5eeb47ef961e65246c, retaining the v0.12 runtime.

The frozen workload was committed before its first formal measurement. Both
formal runs retain the same generator, questions, gold evidence and scorer.
The second run follows diagnosis of the first run's failures: it is a repeated
authored development experiment, not an independent holdout.

Seven checkpoints separate a 3,650-day timestamp shift, other-project growth,
target-history growth, interleaved writes and source replacement. Archives are
built through real ingestion, from 1,024 to 100,000 pages across 788 projects.
Two archived targets stay at 16 pages while six target histories grow. The
optional 200,000-page profile was not run.

Run 44331778554, attempt 1
contains 84 execution blocks, 1,008 quality calls and 12,600 performance calls.
The independent saved-response audit verifies 166,409 checks, all 13,608
responses and 66,136 source references without making new retrieval calls.
The complete original 29,394,152-byte ZIP, exact hashes and a verified restorer
are preserved in Git.

Quality outcome across seven checkpoints Baseline v0.13 candidate
Complete answerable recall configurations 336/392 392/392
Required recall spans recovered 392/448 448/448
Current source first where scored 80/168 112/168
Stale evidence before current 32 0
Complete directory/read configurations 56/56 56/56
Source-coordinate or cross-project errors in saved quality outputs 0 0
Missing-answer queries still returning context 56/56 56/56

At each checkpoint, complete answerable recall rises from 48/56 to 56/56.
All eight multi-source questions now include both the boundary and current
policy. Boundary-first/current-second remains valid and explains why the
current-first metric is 16/24 rather than 24/24. Repeated checkpoints are not
392 independent questions. These measurements concern returned evidence,
not an LLM's final answer or production coding success.

Performance benefits and retained costs

All comparisons below are from the same second run. Six balanced AB/BA rounds
provide 30 cold and 90 warm recall observations per query/arm/checkpoint and
30 complete directory observations. Cold clears the recall-result cache,
not the OS or SQLite page cache. Directory timing includes discovery, section
navigation and original-source reading. Initialization, restore, scoring and
result journaling are outside the query timer.

Pooled p50 direction Faster groups Slower groups
Cold-result-cache recall 12/49 37/49
Warm-result-cache recall 42/49 7/49
Complete directory navigation/read 14/14 0/14
Representative query Baseline p50 Candidate p50
100k interleaved, archived current recall 166.898 ms 102.478 ms
100k interleaved, archived historical recall 133.222 ms 67.912 ms
100k interleaved, archived full directory/read 159.154 ms 26.707 ms
100k grouped, archived full directory/read 157.307 ms 9.958 ms
1,024 pages, archived current cold recall 6.726 ms 10.516 ms
1,024 pages, multi-source cold recall 7.235 ms 11.316 ms
100k interleaved, active current cold recall 332.062 ms 338.309 ms

For the multi-source query, the baseline is incomplete and the candidate is
complete; both timings remain visible. All seven slower warm groups are this multi-source
query, with approximately 6.36–9.33% higher pooled p50. The archived-current
small-corpus slowdown also remains despite both versions returning complete
evidence. The full 112-group review
includes p95, paired-round summaries, resource costs and every adverse group.
The paired-round and pooled statistics are different estimators.
All 672 individual paired records are retained in the
paired-round CSV,
with the complete timing/resource presentation
and its source-bound figures generated without new runtime measurements.

Fixed-size old projects still incur growing cost as unrelated history grows.
For the unchanged archived target, candidate directory p50 increases from
0.7692 to 1.5787 to 4.9298 ms across 1,024, 10,016 and 50,016 total pages.
Global FTS work remains. This release does not achieve constant-time access,
universal speedup, lower memory in every case or a production latency guarantee.
Timestamp age alone preserves all 72 evidence projections per arm; this is
separate from growth and physical-layout effects.

Regression evidence and comparison boundaries

The final equivalent local source repeats the unchanged 800-response
engineering matrix per arm with no score changes or regressions. Book retains
72/72 complete cases, 80/80 spans and 48/48 fixed-policy checks at both budgets
and packing tracks. Each backend preserves 434 positive short-page controls
among 482 configurations, with 48 insufficient-budget controls retained, and
passes 119/119 directory contracts.

Earlier directly bound Python/native diagnostics retain 914 stable-output
pairs, including the 62- and 38-question matrices and 37 independent controls.
Their original source identities remain distinct from the final planner fix.
All rejected attempts, quotation/report counterexamples and both earlier CI
failures remain accessible through the
validation record.

The external-component comparison contains eleven historical pinned
configurations: ten scored configurations plus Memvid's separately diagnosed
query limitation. None was rerun on this 100k workload. Book, BM25S and the
configured Whoosh mode tie on the retained engineering quality counts; Whoosh
uses fewer mean context characters. The 26-family architecture review is not
26 complete products benchmarked. No all-repository superiority is claimed.

Compatibility and delivery

Schema 4, existing source/page IDs, project-bound MCP navigation and the
22-entry native backend remain. The six native C/header files are unchanged.
An existing schema-4 archive requires no new migration. Recall-cache version
26 causes expected cache misses after upgrading without deleting source records.
Defaults remain local lexical retrieval, 2,400 characters and five results;
semantic retrieval and exact-token limits remain optional.

Use the deployment guide
and preserve database paths, project bindings and capture preferences. Restart
clients and workers to load the new runtime. Earlier schemas use the existing
verified-backup and schema-4 migration procedure.

Publication is gated by the exact release source's eleven CI jobs, all thirty
native build/install cells, package checks and the real public-URL installer.
The workflow appends the actual source, CI and release-run links to this body.
The maintained current validation record
records the subsequent independently audited delivery status. Source evidence
alone does not substitute for those delivery gates.

Source commit: 4977ebc13e9da10a8fa4cb95f39e2c533aabe790.

CI: https://github.com/Afloat16/AI-Memory-Book/actions/runs/44343790028

Release workflow: https://github.com/Afloat16/AI-Memory-Book/actions/runs/44343790035