Engrammic記

Wire three agents to a shared memory and the room has no referee. One agent hallucinates that an API uses API keys. A second agent retrieves that claim as context and repeats it. A third builds a plan on top of it. Nobody in that chain ever asked whether the sentence was true the moment it entered the store.

Galileo's analysis of multi-agent failures puts inter-agent misalignment at 36.9% of the total, with context saturation and retrieval noise close behind. Most memory products were built for one agent talking to one user. They have no concept of a claim a second agent should be allowed to doubt.

Engrammic separates what an agent was told from what it has reason to hold. An observation enters as a claim, bound to the source that produced it. Claims accumulate corroboration and heat as other observations point the same direction. A claim that survives consensus crystallizes into a fact.

A stronger claim supersedes a fact rather than overwriting it. A typed edge links the old fact to the one that replaced it, and the old fact stays in the graph, marked as no longer current. Every write also carries two timestamps, one for when the thing happened and one for when the system learned it, so a query for what an agent believed last Tuesday gets a real answer.

$ engrammic write --claim "the API uses API keys" --source session:4f21
✗ rejected: contradiction
  existing fact: "the API uses OAuth2" (crystallized 2026-04-02, 3 sources)
  resolve: supersede the existing fact, or add corroboration to this claim

Model weights are a bad place to keep a belief. They hold roughly 3.6 bits per parameter, shared between generalizing and memorizing, which caps a 70-billion-parameter model at about 31GB of memorization capacity before any of it goes toward reasoning. Correcting a belief stored there means retraining.

Forgetting there follows from the geometry. Gradient descent pushes new-task updates in the directions that damage the old task most, so learning something new degrades what the model already knew as a direct consequence of how it learns anything at all.

Auditing fares no better. Superposition means a single neuron encodes pieces of many unrelated features at once, so asking what a model believes about anything is asking about a mixture, not a fact. Agents writing beliefs into their own separate weights diverge from each other for the same reason distributed systems partition: no shared substrate, no consensus. Brains reached this conclusion first. The hippocampus does fast, sparse encoding of what just happened, and the neocortex does slow consolidation of what usually holds, because one substrate cannot do both jobs at once.

The schema lives in engrammic-primitives, Apache 2.0, defining the four cognitive layers, the edge types, and the scoring functions that decide when a claim earns promotion. The engine sits over a graph store and exposes an MCP server, so any agent that already speaks MCP reads and writes against it without a bespoke SDK.

Manifold, a version of the same engine for latent embeddings instead of text, exists as a design document and nothing else. It waits on a customer who needs multimodal memory.

On 500 annotated coding-agent sessions, the write-gate approach catches contradictions an embedding-only baseline misses: 95% detection against 66%. It propagates a correction through dependent beliefs at 87% against 12% for an append-only store, and blocks 73% of contamination that an append-only store lets straight through. The gate costs latency, about 165ms added at the median, and only runs on the writes that are worth the wait.

An agent that cannot separate what it observed from what it generated does not have beliefs. It has output with confidence attached. Engrammic gives it a way to answer three questions on demand: what does it know, what did it only generate, and which one does a given conclusion rest on.