Maintained Memory

Maintained memory is a design for systems that need to remember more than similar text. It tracks what was observed, what claims are held from those observations, why each claim is held, and what should change when new evidence arrives.

The distinction matters when a fact changes. Suppose an agent records that a person lives in Berlin, then later records that the person moved to Lisbon. A similarity search can retrieve both statements, but similarity alone cannot say which describes the present. A maintained memory represents the reports as evidence, relates them to versioned claims, and computes which claim is current under explicit rules. The Berlin report remains part of the history; it is no longer presented as the current location.

The layers of a maintained memory

The model separates several kinds of data that ordinary text stores often collapse together:

Observations preserve the original reports and their capture context. They answer “what was said or recorded?”
Formation decisions describe what the system extracted or accepted from observations. They answer “what did the processing step propose, and what happened to each proposal?”
Focused memories hold claims, preferences, decisions, events, and other useful statements as immutable versions. They answer “what are the candidate things to remember?”
Contributions and evidence record the acts that support, refine, contradict, derive, or withdraw a claim. They answer “why is this version held?”
Projections are computed views of the canonical records: support, grade, applicability, conflicts, and current alternatives. They answer “what does the evidence justify at this revision?”

These are related but not interchangeable. An observation can exist without producing a focused memory. A stored memory version can exist without being supported. A supported memory can be historical rather than applicable now. Keeping the distinctions explicit prevents a successful write, a model output, or a high search score from being mistaken for a verified belief.

What “belief” means here

In this documentation, a belief is a system-produced label over recorded claims and evidence. It is not a claim that the system has discovered truth. The label is reproducible from the canonical records and the rules in force. Readers can inspect the evidence and derivations that produced it.

The design does not make the model the authority. A model may describe what an observation appears to say, but a deterministic acceptance boundary checks that description against known inputs and rules. Fixed labelling rules then determine whether a claim is reported, corroborated, contested, unsupported, or a hypothesis. Where evidence does not justify a stronger conclusion, the system keeps the weaker one.

That restraint is especially important in a single-user installation. Many reports may all come from the same person and therefore do not constitute independent corroboration. Repetition can make a report easier to retrieve; it does not make the report more trustworthy.

What changes when evidence changes

Accepted writes do not silently edit earlier memory text. A refinement creates a new immutable version tied to its exact parent. A correction, confirmation, derivation, or withdrawal is represented as a contribution with fixed references and provenance. The system recomputes dependent labels from those records.

If one independent justification disappears, a claim may remain supported by another. If the last eligible support is withdrawn, the claim loses support, and claims that depend on it are reconsidered. Claims that only support one another in a circle do not become supported without a grounded source.

Revision is durable and bounded work. While accepted changes are being processed, a current read must not mistake an older projection for the latest answer. The design uses a per-tenant revision barrier: current belief reads are served only when the published revision has caught up with the requested one. If revision is pending, the caller sees that it is pending rather than being given stale labels as current.

A worked example

Consider two observations from the same authenticated source:

“I live in Berlin.”
“I moved to Lisbon last month.”

The first observation remains unchanged. The second is not merely ranked above it because it arrived later. The formation and acceptance rules must recognize it as the same source updating an exclusive location claim, and the report needs enough temporal information to establish when the new state began. If that relationship is established, the system can close the earlier current interval and create a new version for Lisbon. Berlin remains available as historical information.

If another independent source says Berlin, its report does not overwrite the person’s self-report. The claims can be shown as competing evidence, with provenance and their computed standing. If the date of the move is unknown, the system should preserve that uncertainty instead of inventing a precise transition date.

Concepts in this guide

Observations and formation explains capture, evidence identity, automated processing, and the boundary between model proposals and accepted changes.
Beliefs and evidence explains immutable versions, contributions, independent sources, support, grades, and conflict.
Revision and forgetting explains projections, current reads, history, retention, archive, and withdrawal.
Belief-aware recall explains how future recall is planned to represent time, uncertainty, and partial results.
Usage and activation separates content delivered to an agent from content the agent reports using, and explains the planned effect on retrieval order.
Consolidation describes a future way to form broader memories while retaining their evidence and exceptions.
Identity and relationships covers candidate matching, explicit identity decisions, and graph provenance.