Episodes preserve scoped observations while semantic memory stores derived claims

An episode can identify a task, input revision, action ID, tool response, observation time, and externally checked outcome. The tool response may itself be incomplete or untrusted, so recording it does not certify its truth. Preserve enough environment information to know whether the same action would still apply. Several delivery attempts for one tool result are not independent experiences. Likewise, copied reports about one incident remain related evidence even if they appear in different memory rows. Stable source identities prevent duplicate counts from masquerading as corroboration.

A semantic memory record can express a narrower claim such as a procedure working under a particular version and precondition. Include its source episode IDs, derivation version, validity scope, and status. A useful summary does not need every raw detail in the current prompt, but the details should remain retrievable when they matter. If the only stored artifact is a confident sentence with no lineage, later correction becomes difficult: the system cannot tell which observations supported it or which other derived records must be reconsidered.

Reflexion proposes storing verbal feedback from prior trials to influence later behavior without necessarily updating model weights. Generative Agents explores observations, retrieval, and reflection for a different simulated setting. These are concrete research mechanisms with empirical evaluations, not guarantees that generated reflections become factual knowledge. A model can misdiagnose why an attempt failed or infer a broad rule from one accidental success. Keep reflection text distinct from a verified result and preserve the feedback source that produced it.

Compression trades prompt occupancy against detail and future retrieval. A short summary can save query-time tokens while requiring source fetches when an exception matters. It also costs work to create and maintain. Evaluate whether the summary preserves negation, time bounds, artifact versions, and failure cases on tasks that need those details. A high similarity score or repeated retrieval does not make a derived rule universally valid. Contradictions should lead to versioned revision or a scoped alternative, not silent overwriting that leaves an old unsupported conclusion in another index.

Compare raw-episode retrieval, summary-only retrieval, and summaries with source recovery under matched task and compute budgets. Include an outdated procedure, a contradictory later observation, and a case where a rare exception changes the right next action. Measure factual reuse, inappropriate generalization, retrieval count, and externally checked task success. Split related tasks and episodes so the evaluation does not merely replay memorized outcomes. The ability to generate a plausible explanation after failure is different from using a correct source-backed lesson on the next task.

The heat lab illustrates known attenuation, not memory compression. Reset, set both mode n=1 and diffusivity κ=0.5, predict amplitude 1 initially and exp(−0.5)≈0.60653 after two Steps, then inspect. Reset and set both n=4 and κ=0.5; after two Steps predict exp(−8)≈0.00033546. Each Step advances 0.5 seconds in the exact analytic mode exp(−κn²t)sin(nx). The model has no episodes, summary, semantics, or source recovery. It makes loss of fine-scale amplitude explicit while leaving actual summarizer fidelity to separate tests.