memory
Recall with structure
Multi-signal, graph-augmented, fused, explainable — as above. Memory that finds what is connected, remembers what it felt like, and can say why it surfaced.
The suite
Memory is the first organ. The suite is the rest of the body.
RAG bolts retrieval onto a prompt. This is a suite for building agent harnesses — the tools and services a serious agent needs to persist, to act, and to be trusted — designed together so each one strengthens the others.
relationships
Relationship graphs
Not only memory-to-memory. People, agents, projects and artifacts are first-class nodes: who knows whom, who owns what, what depends on what. Recall walks that graph — and so does permission.
world
World models
Memory records what happened. A world model holds what is true now and what to expect next: the state of the environment an agent acts in, its affordances, its rules. Persistent, updatable, and shared between memory and action.
knowledge
Dynamic knowledge bases
Knowledge that revises itself. Contradictions surface instead of coexisting; sources are weighed; stale facts lose confidence rather than lingering as truth. A living graph, not a document dump.
plans
Run plans
Repeatable work as a machine-checkable contract — steps, gates, evidence. A harness executes the plan, checkpoints every transition, resumes after interruption, and escalates when reality disagrees. Reliability by construction, not by re-deriving the job each time.
embodiment
Multi-embodiment
One self, many bodies. The same agent in a terminal, a chat, a voice, a browser, a simulated body — sharing one memory, one relationship graph, one state. Continuity that does not depend on the interface.
worlds
Virtual world exploration
Agents that see and act in virtual spaces through combined visual and textual control — navigate, observe, manipulate, and carry what they find back into memory and the world model.
boundaries
Boundaries that hold
Memory is owned and compartmented. Recall enforces who may see what inside the query itself, with a backstop after fusion — nothing leaks through a path that forgot to check. What an agent is asked to remember is inspected before it becomes a memory; planted text is quarantined, not absorbed.
The aim is not a smarter search box. It is the substrate a serious agent needs to be someone across sessions, bodies and worlds — and the instrumentation its builders need to trust it.
Operated in good faith
Guarantees a memory is held to.
A memory that can be silently rewritten is not a memory. These are the properties the engine is built to hold, whatever the implementation underneath.
- non-erasureWeakened, not deleted
- Content judged unhelpful, outdated or contradicted is down-weighted in a recoverable form — on that judgement, never on a clock. A memory weakened to the floor and left unpinned is pruned only after a grace period. Forgetting is a gradient, never a hole.
- consolidationComposites keep their sources
- When related memories are merged into a composite, the links to the originals are preserved. Consolidation is compression the agent stays in charge of — not a rewrite it cannot inspect.
- contradictionConflicts are recorded, not erased
- When memories disagree, the less-trusted side is weakened and the conflict is written down. Neither side vanishes. Unresolvable or high-severity conflicts are escalated rather than auto-decided.
- isolationRead-time boundaries
- Where a graph spans agents or compartments, isolation is enforced at read time. A foreign memory may influence ranking; it never surfaces another owner's content across a boundary without consent.
- curationInspectable ingress
- What reaches the agent's context is selected by a curation policy that can be read, not a heuristic that cannot. Every recall reports the signals that produced it.
- ledgerAn append-only record
- Retrieval, injection, consolidation, down-weighting, pruning — every operation is logged to an append-only record the agent can query. The history of the memory is itself a memory.
The stack
Boring infrastructure, on purpose.
The full retrieval path runs on PostgreSQL — lexical, affective and structural, with pgvector and pg_trgm. Weaviate carries the semantic path: four named vectors per memory, one per facet. It is an accelerator, not a dependency: switch it off and recall degrades gracefully to lexical matching while the affective and structural paths continue unchanged. Embeddings are pluggable. No proprietary vector cloud sits in the critical path.
The memory service can be co-located with the agent — same host, same network, same operational boundary — PostgreSQL and Weaviate beside it, with no external service in the retrieval path.