菜单
Search and retrieve facts, decisions, and past context from SuperLocalMemory. Use when the user asks to recall, find, search, or "what did we decide/say about X". Triggers multi-channel semantic retrieval with reranking; always call before storing anything new.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
AI agent memory with mathematical foundations. Store, recall, search, and manage memories locally with zero cloud dependency.
Manage SuperLocalMemory session lifecycle — call session_init once at the start of every fresh session to load relevant project context and get a session_id; call close_session when work is meaningfully complete to commit temporal summaries. Correct lifecycle hygiene is what makes SLM's learning loop work.
KV cache for repeated reads — call slm_cache_get(key) first; on a miss do the expensive operation then slm_cache_set(key, value, ttl_seconds) to store it; on a hit use the returned value directly; always fail-open (hit:false on any error, never raises); saves tokens when the same file, query result, or tool output is read more than once in a session.
Compress large text, tool output, or transcripts to reduce context-window usage while keeping the full 1M window intact — call slm_compress(content, mode, reversible, ttl_seconds) to shrink content; if the result is lossy a ccr_id is returned so you can call slm_retrieve(ccr_id) later to recover the exact original; always fail-open (ok:false → continue with the original).
Index and query a codebase as a structural graph — build the code graph, trace blast radius of a change, find callers/callees/inheritors, semantic code search by meaning, assemble PR review context, and detect what changed since last index. Use when the user asks how code connects, what breaks if X changes, what calls a function, what a class inherits from, how to navigate an unfamiliar codebase, or to understand risk before editing.
Capture durable facts, decisions, constraints, and gotchas into SuperLocalMemory. Use when the user says "remember that", "save this decision", "note this constraint", or when a session produces a conclusion worth persisting across sessions. Always recall first to avoid duplicates.
| name | slm-recall |
| description | Search and retrieve facts, decisions, and past context from SuperLocalMemory. Use when the user asks to recall, find, search, or "what did we decide/say about X". Triggers multi-channel semantic retrieval with reranking; always call before storing anything new. |
| when_to_use | - "What did we decide about X?" - "Recall anything about Y" - "Do we have context on the Z feature?" - "Find stored information about authentication / the database / error handling" - "Search for what I said about Y" - Automatically before any non-trivial task, to surface prior context |
| allowed-tools | recall, search, fetch, list_recent, Bash |
Retrieve stored facts, decisions, and past context from SuperLocalMemory using multi-channel retrieval. The golden rule: recall before you remember.
| Situation | Tool |
|---|---|
| Conceptual or paraphrase query ("what did we agree on for auth?") | recall — full multi-channel retrieval + rerank |
| Exact keyword match needed ("find facts containing BM25") | search — FTS5 BM25 only, lower latency |
You have a specific fact_id from a prior result | fetch — exact lookup, full detail |
| Browse newest entries without a query | list_recent |
Use recall as the default. search is a fallback for zero-result recall on a
known exact term. fetch is for when you already know the ID.
Before storing anything new, always call recall first. If a near-duplicate
fact already exists, call update_memory(fact_id, content) to refine it
rather than creating a duplicate. Duplicates degrade retrieval quality for
every future session.
recall(
query="authentication strategy decision",
limit=20, # default 20; reduce to 5 for quick pre-task checks
session_id="<sid>", # pass the session_id returned by session_init
fast=False, # default False; True skips SpreadingActivation channel
)
Real response shape (--json equivalent):
{
"success": true,
"results": [
{
"fact_id": "f8a2bc91",
"content": "Decided to use JWT with 1h expiry for API auth (2026-06-10)",
"score": 0.87,
"confidence": 0.91,
"trust_score": 0.84,
"fact_type": "decision",
"channel_scores": {
"semantic": 0.88,
"lexical": 0.61,
"temporal": 0.72,
"structural": 0.55
}
}
],
"count": 1,
"query_type": "semantic",
"channel_weights": {
"semantic": 0.4,
"lexical": 0.2,
"temporal": 0.2,
"structural": 0.2
},
"retrieval_time_ms": 134,
"no_confident_match": false
}
Always check no_confident_match. When true, no result cleared the
evidence floor. Do not invent a memory — tell the user nothing was found and
offer to search more broadly or store a new fact.
Pass the session_id returned by session_init. It threads engagement signals
through to the ranker so each recall contributes to improving retrieval for
your project over time. Omitting it degrades the learning loop — recall works
correctly, but feedback is not attributed to the session.
Use fast=True for pre-tool-call checks where sub-second response matters.
This skips the SpreadingActivation channel. The remaining channels — semantic,
lexical, temporal, and structural — still run.
recall(query="rate limiting approach", limit=5, session_id="<sid>", fast=True)
When recall returns zero results on a specific term, try search:
search(query="BM25 indexing", limit=10, profile_id="")
profile_id="" uses the active profile. Response has success, results,
and count but no channel_scores or query_type.
fetch(fact_ids="f8a2bc91,d4c1e203")
Returns the full record for each ID: entities, lifecycle, access_count,
importance, observation_date, referenced_date. Use this when the recall
summary (120-char truncation in list_recent) is not enough.
list_recent(limit=20, profile_id="")
Returns facts newest-first. Content is truncated to 120 chars. Use fetch
once you have the fact_id for full content.
recall runs four channels in parallel — semantic vector similarity, lexical
BM25, temporal recency, and structural/graph — then fuses them with Reciprocal
Rank Fusion (RRF) and applies a reranker. The channel_weights field in the
response shows how each channel contributed for that query. Weights adapt over
time based on engagement signals attributed via session_id.
To inspect per-channel scores for a real query against your own data:
slm trace "<query>" [--limit N] [--json]
No benchmark numbers are cited here; performance is workload-dependent.
# Multi-channel semantic recall
slm recall "<query>" [--limit N] [--fast] [--json]
# Opt into shared/global facts for one query (v3.6.15 — off by default)
slm recall "<query>" --include-global --include-shared
# Keyword/FTS5 search (alias: slm search)
slm search "<query>" [--limit N] [--json]
# Per-channel score breakdown
slm trace "<query>" [--limit N] [--json]
# Browse recent memories
slm list [--limit N] [--json]
Flags verified in source (main.py):
slm recall: --limit, --fast, --json, --include-global / --no-global, --include-shared / --no-sharedslm search: --limit, --jsonslm trace: --limit, --jsonslm list: --limit / -n, --jsonMulti-scope (v3.6.15, opt-in): recall is shared-OFF by default — it returns only this profile's facts. Pass
--include-global/--include-shared(or the MCPinclude_global/include_sharedargs) to opt in for a query, or set the defaults in yourmode_a/b/c.jsonconfig. See docs/shared-memory.md.
Flags that do NOT exist (fabricated in old skills — never write these):
--min-score, --format, --project, --tags on recall or search.
If results is empty or no_confident_match is true, report it plainly.
Never construct a response as if a memory was found when it was not. The user
trusts that what you surface came from the store.
SuperLocalMemory v3.6.18 · Qualixar · AGPL-3.0-or-later