Source-backed memory

Pikobot uses Mem0 OSS through instance-local model, embedding and SQLite adapters. No hosted Mem0 account or separate vector server is required. Conversations run through Pi agent-core; memory remains source-backed evidence, not an independent authority system.

What is stored

Background extraction validates facts, verbatim quotes, entities, dates and contextual dependencies, then stores them automatically. There is no memory-confirmation queue. Ask naturally what Pikobot remembers, correct a statement, or ask it to forget a source.

Record	Meaning
Automatically extracted fact	basis: inferred, status: stored; searchable but not verified truth or an instruction.
Explicit canonical assertion	status: current, with its actual owner/document source.
Superseded/history record	Retained lineage for corrections and historical questions, not necessarily current truth.

The human owner, assistant and third parties remain distinct. An assistant question can contextualize an owner's quoted reply; assistant guesses cannot establish owner facts. Only explicit owner preferences become standing instructions, never permissions. Quoted document claims can be deliberately stored with document/revision provenance; document text does not become an owner instruction.

Original statements remain searchable even when extraction misses a detail. Check consequential or conflicting claims against the source. Storage and a matching quote prove provenance, not semantic correctness.

Recall and context

Historical memory/document searches happen on demand. Beyond current dialogue, bounded owner basics, explicit preferences and current attachments, the agent should call memory_search, narrow entity/time filters when needed, then read evidence with memory_read or source_read. Missing startup context does not prove a fact was never supplied.

Source reads can expose bounded related-document references; they are not exhaustive catalogs or the documents' complete contents. Load document tools for canonical reads/search. Installed skills use a metadata catalog and progressive instruction/resource reads, not automatic procedure dumps.

Finite context/tool budgets and model-selected searches can miss evidence or add round-trips. A citation or retrieval hit does not establish that every claim in an answer is supported. Ordinary replies need not print internal IDs; source inspection remains available on request.

Owner basics at chat startup

Authenticated owner foreground/ephemeral chats receive a bounded, query-independent basics block: applicable preferred_name, name, home_city, home_country, occupation evidence. Opening a conversation makes no inference request.

• Only retained owner-authored sources qualify. Inferred values require their verbatim quote; external/assistant and document-confirmation sources are excluded.
• Scope, expiry, source time and explicit-current assertions determine eligibility. Ambiguous alternatives remain labelled, not promoted to verified biography.
• Values/quotes/context are bounded; the total block is at most 2,000 escaped characters and may shrink under context pressure. Unsupported dependencies are omitted.
• Selection makes no model/embedding calls or profile writes. Retained runs link injected evidence for erasure; ephemeral reads create no durable lineage. Restricted background work gets no basics block.

This is fallible evidence, not a permission grant or guaranteed attribution classifier. Forgetting a widely used basic fact can conservatively scrub more dependent replies/instruction snapshots.

Tags and operator inspection

Extraction may suggest a few concise topic tags in its existing call. Tags are optional, lowercase ASCII words/dashes, at most 48 characters and eight per fact. They do not change a fact's basis, source or authority.

The agent's memory_search and operator memory.search support tags, kind, basis, entity, scope and time/history filters. Multiple tags must match the same fact. Eligibility is enforced in SQLite before ranking and rechecked after asynchronous reads; stale vector payloads cannot grant access. Wrong/incomplete tags can miss evidence, so broaden searches when needed.

Authenticated operator RPC examples:

node dist/cli.js --instance PRIVATE_INSTANCE rpc memory.list '{"tags":["food"]}'
node dist/cli.js --instance PRIVATE_INSTANCE rpc memory.search '{"query":"dinner","tags":["food"]}'
node dist/cli.js --instance PRIVATE_INSTANCE rpc memory.read '{"id":"MEMORY_ID"}'
node dist/cli.js --instance PRIVATE_INSTANCE rpc memory.tag '{"id":"MEMORY_ID","tags":["food","health"]}'

memory.tags lists bounded topic counts. memory.remember creates an explicit canonical assertion; memory.retry explicitly requeues failed extraction and may charge. These are administration APIs, not everyday slash commands. Relabelling never changes quotes, timestamps, provenance or embeddings.

Embeddings and implementation

New instances default to DeepSeek Flash chat/extraction and OpenAI text-embedding-3-small embeddings; both can incur charges. Existing settings are preserved. Local E5 or disabled embeddings are explicit alternatives. Missing/failed hosted embeddings degrade visibly to lexical retrieval, not another provider/account. Setup.

src/mem0.ts connects Mem0 to Pikobot's authoritative store and credential services; src/memory.ts validates evidence, namespaces, context and erasure. The pinned dependency patch disables Mem0 telemetry, independent history/default home files and unsafe JSON cleanup for Pikobot's adapter. Dependency upgrades need patch review and regression checks.

Validated source records and Mem0 derivatives commit together after source/cancellation checks. SDK additive extraction cannot silently overwrite explicit assertions. Existing retained records and legacy confirmation lineage survive migrations; no new confirmations are manufactured. Mem0's separate history store is disabled.

Retrieval combines SQLite FTS with a streaming cosine-vector scan, yielding for cancellation. It is O(n), not approximate-nearest-neighbor indexing. Entity/deduplication candidate pools have finite bounds. Large-archive latency, composite/duplicate facts, imperfect keys and omissions remain limitations. The compatibility maxScan setting does not silently truncate the archive.

Forgetting and recovery

Source-aware forgetting removes linked facts, quotes/context dependencies, raw media, vectors, tags and dependent managed code/state/schedules/history. It preserves tombstones against re-extraction and old-backup resurrection. Independent statements are not erased merely because their wording is similar.

SQLite is authoritative; indexes are rebuildable. Backups include memory and lineage in the same snapshot, exclude credentials and repeat erasure checks during restore. Keep the latest deletion ledger and protect/purge external copies. Security · Recovery.

Ephemeral inputs never enter durable extraction. Existing authorized memories may be read, without SDK histories/query caches/entity writes. Providers and OS retention remain separate; volatile does not mean cryptographic erasure.

Verification

pnpm test covers extraction/storage boundaries with synthetic providers, actual SQLite, migrations, cancellation, forgetting and restore. pnpm test:memory:stress uses synthetic dense vectors, not semantic-quality scores. Authorized live memory evaluation is separate and can charge for chat/extraction/embeddings. Follow testing and the evaluation rules; report recall, extraction, attribution, answer accuracy and abstention separately.