Engrammic記

Suppose an agent records "the API uses OAuth" on Monday and "the API uses API keys" on Tuesday. Retrieving the closest text match doesn't tell it which statement applies. The API might have changed, the claims might refer to different endpoints, or one might simply be wrong.

The memory record needs the source, the relevant time, and the reason for a revision.

Engrammic represents those relationships explicitly. An agent can inspect the supporting records and follow revisions instead of relying on the wording of a retrieved sentence.

An observation records an event. A claim adds a statement and its evidence. Promotion rules use confidence and corroboration to decide when a claim becomes a fact; synthesis draws on facts to form beliefs. These are record types and rules for handling evidence, not a guarantee that an accepted statement is true.

The records live as nodes and typed edges in a graph store. A claim points at its source. A fact points at the claims that promoted it. A superseding fact points at the one it replaced. The schema, the edge types, and the scoring functions that decide promotion ship as engrammic-primitives.

The learn operation accepts a claim, its evidence, and its source. Evidence enforcement can reject a claim without evidence or store it with a warning, depending on configuration. A revision can name the record it supersedes. Contradiction checks and later validation help identify claims that need review.

Illustrative records for an API migration:

Earlier claim: "The API uses API keys"
Evidence:      API documentation, version 1

Revised claim: "Version 2 uses OAuth2"
Evidence:      API documentation, version 2
Relationship: supersedes the earlier claim for version 2

Recording a source makes a claim inspectable. Deciding whether the source supports it, and whether two claims actually conflict, still requires validation.

Every write carries two timestamps: one for when the thing happened, one for when the system learned it. These fields distinguish when a statement applied from when it entered the store.

Supersession keeps the earlier record and links it to its replacement. That history lets an agent trace a stored conclusion back to the observations and revisions it depends on. It can only trace what was actually recorded.

I want to inspect and update an individual record without retraining a model. An external store gives each claim an identifier, evidence links, and revision history that an application can query directly.

Source

evidence links

what supports the claim

History

revision links

what changed and what it replaced

Several agents can use the same store. They still need rules for resolving disagreement; sharing a graph alone doesn't make their conclusions consistent.

engrammic-primitives is the schema, Apache 2.0: the layers, the edge types, the promotion scoring. The engine sits over a graph store and exposes an MCP server, so any agent that already speaks MCP can use its tools without a bespoke SDK. Manifold explores a version for latent embeddings.

An internal evaluation note reports the following results from 500 annotated coding-agent sessions, comparing the CITE write gate with a RAG baseline. The dataset and run configuration are not included on this page, so these figures should be read as reported results for that evaluation.

95%

contradictions caught

baseline catches 66%

87%

corrections propagated

baseline reaches 12%

73%

contamination blocked

baseline lets it through

180ms

median write latency

baseline 15ms; increase 165ms

The useful test is whether a correction reaches the next task that depends on it.

The research is at engrammic.ai/research. If you work on agent memory, belief revision, or multi-agent coordination, I would like to compare approaches and evaluation methods.