adaptive-agent-memory-research
A local-first, model-agnostic research repository for long-term agent memory systems. Memory retention is determined by the measured future usefulness of past experience rather than by semantic similarity, using a dual fast-and-slow storage architecture.
Timeline (last 90 days)
The README explicitly states that the append-only log, the controller, and retrieval are built and running, while the measured-future-utility step is still an open research question; the four storage layers below the controller (episodic/semantic/procedural/failure) are design assumptions rather than a validated decomposition.
The repository registers 71 experiments, 37 have been run, 20 are blocked by missing corpora; all are exploratory in nature, and none has been independently reviewed.
The memory-based association graph degraded retrieval, and the authors have retracted their earlier favorable version: its leakage control needed to remove 4.5 edges but removed only 1, causing the gold-standard generation group to reaggregate through two hops.
corpus H1 contains 538 events, 30 simulated days, and 84 delayed probes, spanning seven failure families, deterministically generated, requiring no model or API key; the first run found that document length predicted the answer in all 12 instances across three families.
The project still lacks the reader model (judging whether an arm can answer from what it retained) as well as the compression arm itself; this work is tracked in issue #41, and the README calls it the highest-leverage open task.
The repository is licensed under CC BY 4.0, has no dependencies, requires no API key, and the README states that bugs and rough edges exist, that reports are welcome (see issue #36 for installation problems), and promises to publish regardless of the results.
All Facts (6)
The README explicitly states that the append-only log, the controller, and retrieval are built and running, while the measured-future-utility step is still an open research question; the four storage layers below the controller (episodic/semantic/procedural/failure) are design assumptions rather than a validated decomposition.
50%UnverifiedThe repository registers 71 experiments, 37 have been run, 20 are blocked by missing corpora; all are exploratory in nature, and none has been independently reviewed.
50%Unverifiedcorpus H1 contains 538 events, 30 simulated days, and 84 delayed probes, spanning seven failure families, deterministically generated, requiring no model or API key; the first run found that document length predicted the answer in all 12 instances across three families.
50%UnverifiedThe project still lacks the reader model (judging whether an arm can answer from what it retained) as well as the compression arm itself; this work is tracked in issue #41, and the README calls it the highest-leverage open task.
50%UnverifiedThe repository is licensed under CC BY 4.0, has no dependencies, requires no API key, and the README states that bugs and rough edges exist, that reports are welcome (see issue #36 for installation problems), and promises to publish regardless of the results.
50%UnverifiedThe memory-based association graph degraded retrieval, and the authors have retracted their earlier favorable version: its leakage control needed to remove 4.5 edges but removed only 1, causing the gold-standard generation group to reaggregate through two hops.
50%