local-memory-maintenance

分类: 工具与效率 | 上传者: smarzola | 下载: 0 | 版本: v1.0（最新）
通过提供商工具运行 Hermes 本地内存同行评审、反思、确定性维护、候选人评审、卡片评审和验证。
更新日志: Source: GitHub https://github.com/smarzola/hermes-local-memory
目录结构

当前层级: tree/main/skills/local-memory-maintenance/
📄 SKILL.md 11.2 KB
SKILL.md

---
name: local-memory-maintenance
description: Run Hermes Local Memory peer review, reflection, deterministic maintenance, candidate review, card review, and verification through provider tools.
version: 1.0.0
author: Hermes Local Memory contributors
license: MIT
metadata:
  hermes:
    tags: [hermes, memory, local-memory, maintenance, sqlite, cron]
---

# Local Memory Maintenance

Use this skill when maintaining a Hermes Local Memory SQLite database from inside Hermes Agent or an autonomous cron job.

All flow details live in this skill. Cron prompts should load this skill and provide only deployment-specific parameters such as database path, runtime identity, cadence, and whether to apply or report-only. Do not duplicate the full maintenance flow in cron prompts; update this packaged skill instead.

This skill is intentionally provider-tool-first. Use CLI commands only for non-memory-tool tasks such as creating timestamped file backups, release checks, or deep offline inspection.

## Non-negotiable safety rules

- Never mutate raw messages during maintenance.
- Always make or verify a recent DB backup before applying maintenance changes.
- Prefer provider tools over direct SQLite edits for memory lifecycle changes.
- Treat large, noisy, ambiguous, or identity-confused plans as skipped/escalated, not as work to force through.
- Imported Honcho candidate facts must not be bulk-promoted automatically by deterministic maintenance. They are still valuable first-migration material; review high-signal Honcho memories explicitly and use selected ones to rebuild cards.
- New facts derived from reflection should be candidates first, with evidence message IDs.
- Cards are compact synthesized views. Do not dump every active fact into an existing card.
- Empty-card bootstrap from high-confidence safe active facts is okay when deterministic maintenance proposes it.
- Use `memory_set_card` only for explicit full-card replacement, and never pass an empty card unless intentionally clearing with `allow_empty=true`.

## Canonical provider tools

Read / write cards:

- `memory_get_card`
- `memory_set_card`

Everyday facts and verification:

- `memory_conclude`
- `memory_search`
- `memory_context`

Deterministic consolidation:

- `memory_consolidate` — one subject/observer pair
- `memory_maintenance` — all subject/observer pairs

Packet/review/apply tools:

- `memory_build_peer_review_packet`
- `memory_apply_peer_review_patch`
- `memory_build_reflection_packets`
- `memory_apply_reflection_patch`
- `memory_build_candidate_review_packet`
- `memory_apply_candidate_review_patch`
- `memory_build_card_review_packet`
- `memory_apply_card_review_patch`
- `memory_build_honcho_migration_review_packet`
- `memory_apply_honcho_migration_review_patch`

Legacy names may exist as hidden compatibility aliases, but agents should prefer the canonical names above.

## Full maintenance cycle

1. **Backup**
   - Create a timestamped copy of the SQLite DB before applying changes.
   - If running in Hermes cron, a short Python `shutil.copy2` backup is acceptable.

2. **Peer review / identity**
   - Call `memory_build_peer_review_packet`.
   - If an alias move is obvious and evidence-supported, produce a peer review patch with `alias_moves`.
   - If a runtime/ephemeral peer is clearly the same identity as a canonical peer, produce a peer review patch with `peer_merges`, for example `from_peer_id`, `to_peer_id`, `keep_source_alias=true`, `verified=true`, and a short `reason`.
   - Validate first with `memory_apply_peer_review_patch(apply=false)`.
   - Apply only narrow alias moves or peer merges with `memory_apply_peer_review_patch(apply=true)`.
   - Preserve retired peer IDs as aliases with `keep_source_alias=true` so future lookup and reconciliation remain auditable.
   - Escalate ambiguous identities with concrete peer IDs and aliases.

3. **Reflection / distillation**
   - Call `memory_build_reflection_packets`.
   - For each packet, derive only facts and summaries clearly supported by packet message IDs.
   - Validate with `memory_apply_reflection_patch(apply=false)`.
   - Apply safe patches with `memory_apply_reflection_patch(apply=true)`.
   - New memories from reflection should be candidate facts, not active facts.

4. **Deterministic all-pairs maintenance**
   - Call `memory_maintenance(promote_candidates=true, apply=false)`.
   - Inspect changed-pair summaries.
   - Apply only when the plan is bounded and clearly safe:
     - duplicate candidate supersedes
     - high-confidence local/reflection candidate promotions
     - deterministic empty-card bootstrap from safe active facts
   - If safe, call `memory_maintenance(promote_candidates=true, apply=true)`.
   - Skip large or imported-candidate-heavy plans.

5. **First Honcho migration review**
   - If this is the first adoption/migration from Honcho, do not ignore Honcho candidate memories just because deterministic maintenance will not bulk-promote them.
   - For peers with imported Honcho candidates/cards, call `memory_build_honcho_migration_review_packet`.
   - Inspect the packet's current card, active facts, and candidate facts before drafting a patch. Some peers may already have the useful facts active/carded and only noisy Honcho candidates remaining.
   - Promote only high-signal stable Honcho facts. Prefer not to promote episodic support/chat artifacts, facts phrased around raw numeric peer IDs, tool/system-note artifacts, or one-off medical/logistics questions unless the user has asked to remember them.
   - For noisy imported candidates, either leave them as candidates if unsure or retract them when they are clearly not durable memory material. Supersede duplicates when they overlap with already active/carded facts.
   - Rebuild the compact card from selected imported memories plus existing active facts, keeping it concise and human-readable.
   - Validate with `memory_apply_honcho_migration_review_patch(apply=false)`.
   - Apply safe first-migration patches with `memory_apply_honcho_migration_review_patch(apply=true)`.

6. **Candidate review for noisy imports**
   - For pairs with remaining important candidates, call `memory_build_candidate_review_packet` with a peer and optional source filter.
   - Produce narrow candidate review patches.
   - Validate with `memory_apply_candidate_review_patch(apply=false)`.
   - Apply only selected promotions/supersedes/retractions/card additions.

7. **Card review / synthesis**
   - For sparse, stale, duplicate-heavy, or overly verbose cards, call `memory_build_card_review_packet`.
   - Produce a compact full-card replacement patch.
   - Validate with `memory_apply_card_review_patch(apply=false)`.
   - Apply only if the replacement is compact, evidence-grounded, and preserves important active facts.

8. **Verification**
   - Use `memory_get_card` for affected peers.
   - Use `memory_context` to inspect exactly what prompt injection will contain.
   - Use `memory_search` for key facts that should be retrievable.
   - Re-run `memory_maintenance(promote_candidates=true, apply=false)` and confirm no unexpected residual changes.

## Report format

Report these sections concisely:

- backup path
- peer aliases moved
- unresolved identity prompts
- reflection packets reviewed
- candidate facts added
- session summaries added
- deterministic maintenance changes applied
- Honcho migration review changes applied
- candidate review changes applied
- card review changes applied
- skipped/escalated items and why
- verification results for cards, context injection, and search

## Cron prompt skeleton

Install or update the packaged skill at setup/update time with:

```bash
hermes-local-memory sync-skills --hermes-home ~/.hermes
```

Then attach/load the installed `local-memory-maintenance` skill on the Hermes cron job. Keep the cron prompt short; all flow details live in this skill.

```text
Use the loaded local-memory-maintenance skill to run Hermes Local Memory maintenance.

Database: ~/.hermes/memory/local_memory.sqlite
Mode: apply bounded, policy-safe changes after dry-run validation; skip/escalate ambiguous or large plans.
Runtime identity: use the current Hermes memory tool context. If provider tools are unavailable, instantiate LocalMemoryProvider from the installed hermes-local-memory package with the target database and realistic platform/user/agent identity.

Follow the skill's full maintenance cycle and report format. Do not duplicate the flow here. Never mutate raw messages. Never switch the live Hermes provider config.
```

## Field lessons from cron runs

- If the live Hermes session does not expose the canonical memory tools for the target database/identity, preserve the provider-tool-first contract by instantiating `LocalMemoryProvider` from the checkout/package and calling `handle_tool_call(...)` with canonical tool names. Use terminal/Python only as the wrapper to reach the provider; do not edit SQLite rows directly.
- Initialize the provider with the target database and realistic runtime identity so alias resolution matches live use, for example `LocalMemoryProvider(Path.home()/'.hermes/memory/local_memory.sqlite')` plus `initialize(..., platform='telegram', user_id='151011988', agent_identity='Ambrogio')` in Simone's local setup. This ensures `user` resolves to `simone` and `ai` resolves to `ambrogio` before verification.
- Peer review packets may list many unverified `honcho:*` aliases that are already attached to the intended canonical peers. Do not move aliases just to mark them verified. Apply `alias_moves` only when a source alias clearly points at the wrong peer. Apply `peer_merges` only when a runtime/ephemeral peer is clearly the same identity as a canonical peer; set `keep_source_alias=true` so names such as `telegram-default` continue resolving after the stale peer row is removed. Use `human_prompts` when the safe action is unclear.
- Reflection packets can mix durable preferences with stale operational, task-local, or sensitive imported history. Apply reflection patches only for windows with clear evidence IDs and stable memories; skip old mixed Honcho windows rather than forcing a summary/fact extraction. After applying reflection summaries, rerun `memory_build_reflection_packets` to confirm only intentionally skipped windows remain.
- Deterministic maintenance after reflection may be safe when proposed promotions are all high-confidence local/`agent-reflection` candidates and counts are small. It remains unsafe for broad imported-Honcho promotion. Record the exact dry-run counts before apply.
- Candidate review is useful for narrow cleanup of noisy imported Honcho facts. Retraction is appropriate for obvious task-local or ephemeral candidates (for example one-off light-control commands or “should this be saved?” meta-candidates); do not promote imported candidates in bulk.
- Deterministic maintenance may append newly promoted safe candidate lines verbatim to cards. Follow with card review when this makes a card too verbose or task-specific, and replace the whole card with a compact synthesized version rather than dumping every active fact into it.
- Final verification should include `memory_get_card`, `memory_context`, `memory_search`, and a final `memory_maintenance(promote_candidates=true, apply=false, limit=...)`. A clean final dry-run should report zero unexpected promotions/supersedes/card additions; any remaining reflection packet should be explicitly reported as skipped/escalated.
登录后下载/点赞/收藏 ❤ 17 | ★ 0
local-memory-maintenance

目录结构

SKILL.md

举报内容

提示
