A database returns what was written. A memory engine reconstructs what is reachable from a cue at this moment — under decay, association, and tier physics — without losing a live trace because the first index was the wrong one.
This repository contains the Memory Fundamentals specification — a formal model, algebra, and conformance rules for memory engines, in the tradition of Codd's relational model for databases.
The industry classifies "agent memory" by storage topology: vector indexes, graph stores, document databases, orchestration layers. Those are storage patterns. None is a definition of memory.
This specification draws the distinction. It defines:
- A data model — traces, cues, signals, associations, tiers, and strength as first-class concepts
- A recall algebra — closed operations over memory state (remember, recall, rehearse, reconsolidate, forget, consolidate, simulate, dream, and more)
- Normal forms — integrity constraints a well-formed memory must satisfy (NF0–NF7)
- Twelve conformance rules — what an implementation must demonstrate to be called a memory engine (M0–M12)
The identity of a memory engine is its recall algebra, not its storage topology.
| Document | Description |
|---|---|
| MF-001 | A Memory Model of Data for Persistent Agents — Model, Algebra, and Conformance Rules for Memory Engines |
A plan that takes cosine top-k then re-ranks by importance will often miss the trace that matters most — an old, high-importance, emotionally vivid memory with low cosine similarity to the current cue. This is the truncation trap. Rule M2 forbids single-signal candidate generation as the sole generator.
Databases have existence. Memory has two axes: retrieval strength D (current accessibility, decays with disuse) and storage strength S (encoding durability, non-decreasing). Forgetting is loss of accessibility, not loss of storage. High S slows future loss of D — the protective effect of strong encoding (Bjork & Bjork, 1992).
Every trace carries a closed provenance tag: experienced, distilled, simulated, or rehearsed. Offline simulate and dream operations may generate hypotheses; they must not mint autobiographical episodes. Default recall hard-gates simulated traces.
Spector is cited as an existence proof that the model can be implemented on a single substrate. It is the companion, not the definition.
Spectrayan. 2026. "A Memory Model of Data for Persistent Agents."
Memory Fundamentals Specification MF-001, v1.0.0. August 2026.
https://github.com/spectrayan/memory-fundamentals
- Model, algebra, and twelve rules (MF-001 v1.0.0)
- RCL (Recall Language) BNF grammar
- Conformance test suite
- M8 lineage record format
- Truncation-trap benchmark corpus
This specification builds on and acknowledges:
- ACT-R (Anderson et al., 1993–) — formal cognitive architecture; the closest ancestor
- Soar (Laird, Newell, Rosenbloom, 1987–) — production-system cognitive architecture
- Bjork & Bjork (1992–) — new theory of disuse; two-factor strength model
- CoALA (Sumers et al., 2023–24) — cognitive architecture taxonomy for LLM agents
- Animesis (Li, 2026) — constitutional memory architecture; governance axioms
ACT-R and Soar formalize how cognition works; this specification formalizes what an engine must do. The distinction is between a theory of mind and a specification for a machine.
This work is licensed under CC BY 4.0.
Comments and counterexamples are invited. A specification that cannot be challenged in public is not a specification.