Category 03 of 12
The Brain
How Agentis agents form, recall, and ground memory over time: the formation pipeline, hybrid recall with citations, knowledge bases, and living skills.
Pages in this category
All Brain memory lives in one table, discriminated by plane, scope, and type — why a single substrate beats a dozen disconnected memory stores.
Writing memory is reconciliation, not append: deterministic extraction, a two-phase Formation Judge (ADD/UPDATE/NOOP), PACER routing, and Feynman repair.
The real Feynman technique made concrete: force a compact explanation of a failure, validate it against real evidence, and commit nothing at all if it isn't actually grounded.
Hybrid TF-IDF plus embedding recall with freshness decay, trust weighting, MMR diversification, and citation-or-abstain grounded answers.
An evidence ledger, versioned sources, claims with an approval workflow, conflict resolution, and per-agent grants — the Brain's organizational-intelligence engine.
Ingest, chunk, embed, auto-link, and enrich uploaded documents — PDFs, docx, and text — into searchable knowledge atoms.
Body on disk, metabolism in the Brain: a real SKILL.md a harness loads natively, paired with a confidence-tracked skill atom that graduates by reuse.
Private, per-user notes and a personal knowledge graph, distinct from the shared workspace Brain.
The cross-scope brain graph that unifies chat-as-dispatch with the rest of workspace memory.
Immediate-if-overdue and daily lifecycle, quality hygiene, compression, reembedding, and health checks that keep recall honest.
The scoring mechanics behind recall: freshness decay, trust weighting, reranking, and MMR diversification, explained with real inputs.
A LoCoMo/BEAM-style eval harness runs as a CI gate on every change to the Brain — single-hop, multi-hop, temporal, knowledge-update, unanswerable, and mid-task recovery, measured, not assumed.
Why deciding what to remember is a judgement call, what the on-device model can and cannot do — measured — and the tier you choose in Settings.