Beyond Context Windows: Evaluating Long-Term Memory for AI Agents
Frames evaluation of long-term agent memory, a gap retrieval-only tests leave open.
AI Storyline
3 items · 2 sources · 3 days
Operational story trace
Long-Term Memory
Follow in this browser to see new updates on your Live feed.
Current stateDeveloping
Latest change
An open-source repo (Oct 11) stores coding-agent long-term memory as a tree of plain files, a simple alternative to a dedicated memory service.
Earlier contextThe story so far
A Sep 22 post asked how to evaluate agent memory beyond the context window. PACMI (Oct 5) then shifted the focus to keeping memory correct by invalidating stale entries.
editor-curated · source-linked
Arc
Sep 22Oct 11 · now
EVALUATION · Sep 22A benchmark post asks how to evaluate memory beyond context windows1 source · show source ▾
CORRECTNESS · Oct 5PACMI cascades invalidation through provenance-linked memories1 source · show source ▾
PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents
Adds provenance-aware invalidation for memories that go stale.
NOW · Oct 11Coding-agent memory as a tree of plain files1 source · show source ▾
Long-term memory for AI coding agents as a tree of plain files
A file-tree implementation of long-term memory for coding agents.
Day 1 Tuesday, Sep 22, 2026
Beyond Context Windows: Evaluating Long-Term Memory for AI Agents
Frames evaluation of long-term agent memory, a gap retrieval-only tests leave open.
Day 2 Monday, Oct 5, 2026
PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents
Adds provenance-aware invalidation for memories that go stale.
Day 3 Sunday, Oct 11, 2026
Long-term memory for AI coding agents as a tree of plain files
A file-tree implementation of long-term memory for coding agents.
What to watch — open questions
- Does PACMI's invalidation cascade hold up on an independent long-horizon agent benchmark?
- Will any file-based memory store publish results on stale-memory handling?
How this thread was built
editor wrote the arc · 3 beatswatcher 1 status change
Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.