v14.7.0

Whole records fill the context budget, and TAM on the AMA-Bench board.

Recall

• memory_recall(mode="context", fill_budget=true) searches up to 100 hits deep and keeps whole records, in rank order, until context_max_chars is used. The plain context mode excerpts the top limit hits, so with short records most of the budget stays empty. Off by default.
• Context that holds the gold evidence: LoCoMo (383 questions, 8,000 characters) 0.815 → 0.890; LongMemEval-S (94 questions, every gold turn, 40,000 characters) 0.766 → 0.904, multi-session questions 0.333 → 0.733. Details: context-fill-v14/RESULTS.md.

Benchmarks

• Adapters for the official AMA-Bench and LongMemEval-V2 harnesses. Paid calls go through a local proxy that holds the OpenAI key and stops the run before a dollar ceiling.
• AMA-Bench, full test set (208 episodes, 2,496 questions), TAM as the memory of a gpt-5-mini agent: 0.658 (mean of the 24 domain × capability cells, judge gpt-5.2). GPT-5 mini reading the whole trajectory scores 0.656 on the published board. Submitted to the leaderboard; self-reported until the board's judge verifies it. Configuration and command: ama-bench/README.md.

Fixed

• A code record saved while the code embedding model could not load (offline, or after macOS purged the model cache) was labelled with the code model and became invisible to semantic search. It is now stored in the text space.
• Credential redaction no longer breaks UUIDs whose first groups are all digits (a team-server memory_save could fail with "request_id: Input should be a valid UUID").

Every change: CHANGELOG.md.

Install: pip install total-agent-memory==14.7.0 · npx -y total-agent-memory connect <ide> · brew install vbcherepanov/tap/total-memory · docker run -p 37737:37737 -v ~/.tam:/data ghcr.io/vbcherepanov/total-agent-memory:14.7.0