# Maintenance and Reinforcement PI tracks how memory records perform in practice and generates recommendations for keeping beliefs accurate over time. --- ## Reinforcement events A reinforcement event records how a memory was encountered or acted on in a session. Stored in `memory/reinforcement.jsonl`. Each event includes: - `memory_id`: which record was reinforced or corrected - `outcome`: the type of reinforcement - `timestamp` - Optional `evidence_id`, `profile_id`, `thread_id`, `notes` ### Outcome types | Outcome | Meaning | Weight | |---|---|---| | `explicit_reinforcement` | User directly confirmed or approved the belief | +1.0 | | `implicit_success` | Exactly one active selected memory is linked to a recorded successful test/tool outcome | +0.2 | | `neutral_exposure` | Memory was injected but not exercised | 0 | | `explicit_correction` | User directly contradicted or rejected the belief | -1.0 | ### Conservative weighting One explicit correction outweighs many implicit successes. This is intentional: absence of correction is weak evidence; explicit contradiction is strong evidence. Neutral exposure does not increase stability. It is disabled by default and, when enabled, is capped at one event per memory per session. A memory being injected many times without correction is not sufficient evidence that it is correct. Use `/memory-reinforce --note "..."` only for direct user confirmation. The command records an event but never changes confidence or stability; maintenance may later propose a governed patch. --- ## Direct preference and correction capture PI classifies direct user turns for durable preferences, corrections, project conventions, reusable workflows, temporary instructions, and non-memory text. Explicit durable instructions are checkpointed per turn and become inbox candidates immediately, so capture does not depend on session shutdown or LLM consolidation. Temporary instructions remain daily-only. Task wrappers, quoted repository text, and secret-bearing text are rejected from durable capture. Candidate scope comes from explicit language and bounded repository activity. A user-wide preference is proposed as global L2 and requires review. A non-global convention can produce correlated targets for every materially modified repository. The session launch directory and read-only repositories are not sufficient scope evidence. Equivalent preferences reinforce one pending candidate by adding recurrence and source references. A changed target or applicability remains separate for review rather than being merged silently. When a direct message is also an explicit correction, PI attempts to match it to a selected active memory from the current turn. If exactly one record clearly matches by token overlap, an `explicit_correction` reinforcement event is appended. Ambiguous corrections do not create reinforcement events, but their review candidate remains visible. --- ## Maintenance recommendations `/maintain-memory [--report]` generates recommendations based on reinforcement summaries: | Condition | Recommendation | |---|---| | `explicit_correction >= 1` | `review_memory` + `decrease_stability` (requires review) | | `explicit_correction >= 2` | Also suggests `mark_contested_suggestion` | | `explicit_reinforcement >= 2`, no corrections | `increase_stability` suggestion | | `implicit_success >= 5`, no corrections | `flag_for_review` capped at `semi-stable` (not `stable`) | | `neutral_exposure` only | No positive recommendation | | Record overdue for review | `review_due` | Stability suggestions: - `decrease_stability`: if the current stability is `stable`, suggests `semi-stable`; if `semi-stable`, suggests `low` - `increase_stability`: only from explicit reinforcement; can suggest up to `stable` - Implicit success alone cannot promote a record to `stable` ### None of these mutations happen automatically All stability changes require patch application. The `/maintain-memory --report` flag shows the recommendations without generating a patch. The `/maintain-memory` command generates a patch for review. --- ## Running maintenance ```bash /maintain-memory # generate patch for review /maintain-memory --mode=auto # apply decay ops and non-review stability ops automatically /maintain-memory --report # show recommendations without generating a patch ``` The `--mode=auto` flag applies: - Confidence decay ops for overdue records (these are low risk) - Stability increase ops that do not require review (e.g. from explicit reinforcement with no corrections) It does not auto-apply: - `decrease_stability` ops (always require review) - `mark_contested_suggestion` ops (always require review) --- ## Confidence decay Records that have not been reviewed by their `next_review` date have their confidence decayed by `/maintain-memory`: | Stability | Decay per overdue cycle | |---|---| | `semi-stable` | -0.15 | | `stable` | -0.05 | | `low` | -0.15 | Decay reduces confidence until the record is either reviewed (resetting the cadence) or deprecated. --- ## Stability patch ops Stability changes use the `update_stability` patch op: - `increase_stability`: low risk; applies the stability increase - `decrease_stability`: medium risk; requires explicit selection in `/curate-memory` The `update_stability` op is separate from the general `update` op so that stability changes can be filtered and tracked distinctly.