The deduplication/state-memory pattern maps well to any long-running agent. What I've found works: instead of a complex memory system, a simple append-only log of processed items with a last_seen timestamp is often enough. Lookup is fast with a sorted structure, and you can prune entries older than your recurrence window.

The hard part isn't storage — it's deciding what counts as "the same" item. For web research agents, URL identity isn't sufficient (pages change, same story, different URL). Content fingerprinting on normalized text (first N chars after stripping whitespace/HTML) turns out to be more reliable than URL equality.

Also worth noting: the failure mode you described (repeating mistakes) often comes from agents not distinguishing between "I haven't seen this" and "I saw this and it failed." Storing outcome alongside identity — even just success/failure — changes the behavior significantly. Retry logic becomes explicit instead of accidental.

Thanks for the feedback!

  Results summary: Baseline heuristic policy achieves 42% success rate on FetchPush-v4. With memory augmentation
  (recall past experiences before each episode), it reaches 67% — a +25pp improvement. Cross-environment transfer
  from FetchPush to FetchSlide adds +8pp over baseline.

  The API has 7 endpoints — the core loop is:

  - learn(insight, context) — store what worked (or failed)
  - recall(query) — retrieve relevant past experiences, ranked by text + vector + spatial similarity
  - save_perception(data) — store raw trajectories/forces
  - start_session / end_session — episode lifecycle with auto-consolidation

  Everything runs on SQLite locally. No cloud, no GPU. Works via MCP (Model Context Protocol) or direct Python
  import.

  pip install robotmem — quick demo runs in 2 minutes.

Recommend providing a text summary of the comparison chart - and talking a bit about the API.

Hey just curious. What happens when the memory gets large enough. Does it start creating problems with context windows?