Paper

Reverie: A Biological Memory Architecture for AI Agents

M. Sarda5 September 2026Version TODO(fact)

This page is the abstract, the corrections and the citation. The full text will appear here once the paper’s corrections are made.

Cite

Abstract

TODO(copy) The paper has no abstract; the owner writes about 120 words.

What the paper covers

The paper has 17 sections. Its biology (complementary learning systems, reconsolidation, the forgetting curve and synaptic homeostasis) is presented as inspiration and convergence, with citations.

  • §2, §15Extraction of conclusions.
  • §3What is kept: conclusions, not facts or code descriptions.
  • §4, §8Confidence and decay.
  • §5Two brains and the review gate (Fig. 4).
  • §6Maintenance and contradictions.
  • §7Retrieval and ranking.
  • §9Evidence and the edges between conclusions.
  • §12Benchmark runs, with figures that do not agree with our other write-ups (see EC-Bench).
  • §14–§16Vision material, which the paper itself frames as speculative.
  • §17The MCP interface.

Sections not listed: TODO(fact) The plan does not state their contents or the section titles.

Where the paper and Reverie differ

The paper was written on 5 September 2026 and has not yet been corrected. Twelve passages describe something Reverie does not do, or do it differently. Where they differ, the right-hand column is what Reverie does.

  1. 01The agent’s tools

    The paper says
    The MCP server exposes ec_query, ec_start, ec_stop and ec_status.
    Reverie does
    The four tools are ec_observe, ec_query, ec_get_summary and ec_reconsolidate. Starting, stopping and checking status are commands you run in a terminal.

  2. 02The extraction model

    The paper says
    Extraction uses a specific named model.
    Reverie does
    Extraction uses whichever model is configured, hosted by default or Ollama for local use. This site names no model.

  3. 03Retrieval diversity

    The paper says
    Retrieval applies maximal marginal relevance.
    Reverie does
    It does not. Results are ranked with a four-factor score, and duplicates across groups are removed.

  4. 04Ranking factors

    The paper says
    Ranking uses similarity, confidence, scope match and retrieval frequency.
    Reverie does
    The score combines similarity, effective confidence, activation and network richness, then is multiplied by trust and by scope proximity.

  5. 05Confidence bumps

    The paper says
    Retrieving a conclusion also raises the confidence of the conclusions connected to it.
    Reverie does
    Activation is a separate ranking score, kept per session. The confidence bump applies only to the reviewed conclusions that are actually surfaced.

  6. 06Who creates edges

    The paper says
    The extractor creates depends_on edges.
    Reverie does
    The extractor never creates edges. A separate step that relates new conclusions to existing ones does.

  7. 07Maintenance tasks

    The paper says
    Maintenance applies decay and detects contradictions.
    Reverie does
    Decay is computed when a conclusion is read and is never written back. There is no standalone contradiction scan. Maintenance does forgetting, grounding checks, edge pruning and clustering.

  8. 08Contradictions

    The paper says
    Both contradicting conclusions are marked challenged.
    Reverie does
    The existing conclusion is marked challenged. Both sides become open questions only after a persistence limit that depends on scope.

  9. 09Rejected evidence

    The paper says
    If evidence is rejected at review, its effect is subtracted and the edge remains.
    Reverie does
    Evidence from a session never changes a reviewed conclusion before review, so rejected evidence is discarded. Edge pruning later reverses the effects of evidence from conclusions that have been retired.

  10. 10“Nothing crosses without review”

    The paper says
    Nothing crosses from session memory to long-term memory without human review.
    Reverie does
    True for conclusions extracted from sessions. Updating an already reviewed conclusion through ec_reconsolidate takes effect immediately, and its supersede path can create a new long-term conclusion. This is a documented tension.

  11. 11Half-lives

    The paper says
    Decay has a half-life of a given number of days.
    Reverie does
    Those figures are half-lives of the odds (the log-odds), not of the confidence value. Read them that way if they are reproduced.

  12. 12A traced-requirements count

    The paper says
    The paper states a count of requirements traced.
    Reverie does
    We cannot verify the count against the implementation, so this site does not repeat it.

Cite

@misc{sarda2026reverie,
  title = {Reverie: A Biological Memory Architecture for AI Agents},
  author = {Sarda, M.},
  year = {2026},
  month = sep,
  url = {https://reverie.muditsarda.com/research/biological-memory-architecture}
}

What this changed in Reverie

How it works describes what Reverie does today, including the points above where it differs from the paper.

What this changed in Reverie