Memory that can explain its sources

An agent that recalls a sentence from a database has retrieved information. It has not necessarily learned in the sense of changing model parameters.

A useful memory architecture distinguishes these mechanisms rather than placing them under one impressive name.

Define what persists

A system may preserve raw documents, extracted facts, conversation summaries, task state, embeddings, or model updates.

Each has different costs and failure modes. A summary can omit a qualification. A fact can become stale. An embedding can retrieve a similar but irrelevant passage. A model update can be difficult to reverse or attribute.

State which kinds of persistence the application actually implements.

Preserve provenance with the claim

A memory record should identify its source, creation time, relevant subject, access scope, and any revision or expiration condition.

Separate an observation from an inference. “The test passed on this revision” is different from “this method always works.”

When facts conflict, retain the evidence needed to resolve them. Silently replacing an old statement with a new confident summary erases useful history.

Use semantics where semantics are needed

Hashes are useful for exact identity and integrity. Semantic embeddings are learned representations intended to support different similarity relationships.

Neither guarantees truth. A semantically close document can still be wrong, unauthorized, or out of date.

Evaluate retrieval with questions whose relevant sources are known. Measure whether the correct passage appears and whether the final answer uses it appropriately.

Keep access checks in retrieval

Do not retrieve private content and hope the model will avoid revealing it. Filter or authorize at the data boundary.

The scope of a stored memory can change when a project membership changes. Test that revocation affects future retrieval, exports, and cached results.

A “private” label without enforcement is metadata, not protection.

Measure the effect on the task

Compare the same task with and without the memory feature. Include cases where old information is misleading.

A larger knowledge graph can increase available context and also increase noise. Node count is not a capability metric.

Track source accuracy, task completion, stale-answer rate, retrieval cost, and the ability to correct or remove a record.

Evaluate procedures as well as remembered facts

The May 2026 LongMemEval-V2 preprint evaluates memory across long task histories, including changing state, workflows, and pitfalls. The important extension is that remembering what happened is different from using the right procedure on a later task.

Create paired tasks: one establishes a fact or procedure, and a later one requires using it under slightly changed conditions. Include corrections and expired facts. Score successful task completion, correct source use, and appropriate refusal to reuse obsolete information separately.

A remembered command should retain its environment assumptions. A successful database migration from one repository is not automatically a valid procedure in another.

Consolidate with a correction path

The August 2026 Lychee Memory V2 preprint describes semantic segmentation of conversations and typed, context-independent memories with indexed retrieval. Its reported efficiency and benchmark gains belong to the authors’ experimental setup.

A practical implementation can trial a small set of memory types: observation, user preference, decision, procedure, and unresolved claim. Each entry carries its source, effective time, scope, and supersession links. Consolidation should retain disagreements that matter rather than producing a more confident average.

Compare three configurations under the same tasks and budget: no persistent memory, simple retrieval of source excerpts, and the proposed consolidated store. Measure construction and retrieval cost as well as outcome quality. Inspect cases where memory makes performance worse.

Finally test correction propagation: update the source, invalidate or supersede its derived memory, and ask the dependent question again. The system should expose the current evidence and preserve the older claim as history, not keep repeating it because it was retrieved often.

Reserve learning claims for learning

The Titans research concerns a neural memory mechanism. A local cache inspired by its structure should be described as a cache until it implements and tests a comparable learning process.

The distinction does not diminish ordinary retrieval. A carefully maintained source collection can be extremely useful.

The strongest memory is one that helps the system answer correctly and can show why a remembered claim deserves confidence.