feat(reconsolidation): mutate memories on retrieval, gated against distortion

[nickroci]

What

Implements the reconsolidation pathway from the README roadmap: retrieved memories become labile and are re-stored mutated. Adds a drift salience signal so that when an entry is genuinely used this turn (the Tier-2 / used_helpfully moment), the Librarian may reconsolidate it.

The design is built to harness the update half of reconsolidation while defending against the distortion half — Bridge & Paller (2012), "every retrieval is a partial re-write that measurably distorts." (Now cited in the README references.)

How it works

• Trigger — an entry surfaces and is used this turn → the Librarian reads its full body (read_entry returns it untruncated) and may propose a drift update_entry.
• Three outcomes — integrate a genuine new qualifier from the use-context; sharpen the entry; or, if it has grown mixed-topic, split it into linked entries (reactive curation, not aggressive grooming).
• Guardrails against distortion:
  1. Claim-preservation gate — the Scholar diffs old-vs-new and vetoes any change that weakens the load-bearing claim or is mere churn.
  2. reconsolidated / last_reconsolidated counter — tracks cumulative mutation so the Scholar tightens on much-churned entries. Preserved across ordinary rewrites; incremented + stamped on drift updates.
  3. Split-on-sprawl, not a char cap — size is managed structurally by the Librarian's reactive split. Growth only happens on use, and use is the same moment the curator reads the entry and can decompose it, so an entry can't silently balloon.

Design notes

An earlier iteration used a deterministic non-growth size cap; it was removed because (a) the corpus already holds 24 KB entries, so a flat cap would brick the biggest ones, and (b) growth and split-opportunity share the same use-trigger, so the curator-driven split is sufficient. The residual risk is prompt-quality ("does the split instruction fire reliably"), which is tunable and observable via the counter. An optional deterministic size→split backstop is left as a documented TODO.

Files

• _schemas.py — drift added to the salience-signal enum
• librarian_prompt.py — drift classification (integrate / sharpen / split-on-sprawl)
• scholar_prompt.py — drift review block (claim-preservation, split-not-cram)
• scholar_executor.py — reconsolidation counter bookkeeping
• tests/test_scholar_executor.py — counter tests
• README.md — roadmap row flipped to shipped + Bridge & Paller reference; stale double-install quick-start blockquotes removed

Verification

ruff + pyright (source) clean, duplicate-code gate passes, full daemon suite 662 passed, coverage 90.45%.

🤖 Generated with Claude Code