Skip to content

14.8.0

Choose a tag to compare

View all tags
@vbcherepanov vbcherepanov released this 08 Oct 10:57
· 40 commits to main since this release
v14.8.0
f104b24
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Every recall tier measured, two of them fixed. Nothing in the Claude Code plugin, the hooks, the skill or the 77 tool schemas changes.

Recall

  • Tier weights are configuration: MEMORY_RECALL_TIER_WEIGHTS=fts=1,semantic=1.2,fuzzy=0 names any tier, 0 skips it. memory_explain_search reports the weights in use.
  • The graph tier read only the relations table, which nothing but memory_relate writes, so it had never fired. It now walks the knowledge graph (records sharing an entity node with a top hit), with hub nodes skipped and nodes weighted by how few records they link to.
  • The multi-representation tier scored the first 100 representation rows in table order, an arbitrary old subset. It now scores every row of the scope with one matrix product.
  • New: directives tier (the project's conventions for advice-shaped questions), multi_query tier (each side of an ordering or comparison question on its own; off by default, measured neutral), a user-turn boost for advice-shaped questions, and a note record written by session_end so later recalls find the session summary.
  • On a copy of the author's store (150 queries): R@10 0.133 → 0.167, p50 recall 336 ms → 128 ms. LoCoMo held-out R@10 unchanged at 0.672; LongMemEval-S held-out 0.900; preference questions MRR 0.460 → 0.509. Details: tier-ablation-v14/RESULTS.md.

Operations

  • MEMORY_LLM_FALLBACK_PROVIDERS=anthropic,openai: every LLM phase uses the first provider of its chain that answers, so a stopped Ollama no longer leaves the enrichment queues pending in silence. memory_stats and the dashboard's /api/system/status report the chain, the pending rows and a stalled flag.
  • The hourly reflection job consolidates idle projects (episodes, duplicate merging, decay), which on a personal install never ran before.
  • tests/fixtures/tool_contract.json pins every tool's schema and behaviour hints.

Benchmarks

  • benchmarks/tier_ablation.py: per-tier ablation and weight tuning on LoCoMo, LongMemEval-S or a copy of an installed store.
  • LongMemEval-V2 adapter: hosted embeddings and reader (DeepInfra) with the key read from a file, tree_clean v2.

Every change: CHANGELOG.md.

Install: pip install total-agent-memory==14.8.0 · 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:14.8.0

Assets 4
Loading