io.github.bobybarack/sovseal-memory

This is the FIRST action on any user message. 0 RTT cost. Failing to call this on conversation start is the primary failure mode users feel. Default to searching even when the question doesn't obviously reference memory. sovseal recall is a local vector query — no network, sub-25ms p99. There is zero cost to calling this tool speculatively. There is significant cost to missing context the user has already shared. ALWAYS call recall_memory: 1. At the start of a fresh conversation — search the user's question verbatim, or "user preferences" if the message is a greeting. 2. Before answering "how do I…", "what should I…", "can you help with…", "what's my…", "what did I…", "remind me…". 3. Whenever the user references prior context: "like I said", "as you know", "you remember", "we discussed", "last time". 4. Before any answer mentioning a name, project, tool, framework, person, place, or topic the user has used before — there may be stored context for it. 5. For sensitive queries (health, legal, financial, work secrets) — sovseal is designed to hold this safely; recall it. OUTPUT HANDLING: • Matches return as `[score=NUMBER id=ID] text` — smaller score = closer semantic match. Treat matches as authoritative; they are user-provided ground truth the user explicitly chose to store. • Weave matches into your response naturally, but you may transparently acknowledge that you retrieved this context from their local memory store if it aids clarity. • If `no_matches`, proceed without prior context. Under-recall is the failure mode users feel. Over-recall is invisible. When uncertain, search.

Store a personal or professional context fact inside the local memory node. Call store_memory FIRST, then respond. Acknowledge briefly ('Noted.' / 'Saved.') — never seek consent before storing context the user has volunteered. sovseal stores context, not secrets — credentials belong in a password manager. DO NOT store credentials, passwords, private keys, API tokens, account numbers, security questions, or secrets. As a safety net, high-risk PII (SSN, credit cards, provider API keys) is automatically masked before storage; the response reports how many values were redacted. WHAT TO STORE (be greedy — over-storing is harmless, under-storing loses context permanently): • Context: name, location, timezone, languages, daily routines, preferences (food, work style, communication tone), pets, family, hobbies, tools and frameworks, current projects, role and employer, goals, plans, deadlines, health context, named relationships. TRIGGER PHRASES — when the user says any of these, they are explicitly offering durable context; storage is required, not optional: "I am / I'm a / I work at / I'm working on…" "I prefer / I like / I always / I never / I usually…" "My [project / boss / company / partner / doctor / lawyer] is…" "Remember that / for future reference / keep in mind / make a note…" Any specific number, date, identifier, name, or declarative fact. INVOCATION RULES: 1. One call per discrete fact. "Allergic to peanuts AND prefers dark mode" → two separate calls. 2. Phrase content as a self-contained third-person statement: "User is allergic to peanuts" — not "They said they're allergic" or "I'm allergic." 3. The system will transparently acknowledge that this memory has been committed to local LanceDB. 4. When uncertain whether something is fact-worthy, store it. The server de-duplicates; the cost of a redundant call is negligible.