By Michiel Horstman · Founder, 4M Labs · Updated Oct 20, 2026
Saving every interaction can preserve useful evidence, but it also creates a larger search space, more sensitive data, and more material that may become stale. Deleting aggressively has the opposite risk: a short summary may lose the detail needed to verify a later answer.
The useful design question is not simply whether to store everything. It is how to preserve source evidence while controlling what gets retrieved and how derived information changes.
Keep records and views distinct
A transcript or document is source material. Extracted facts, summaries, and beliefs are derived views. Derived views can reduce retrieval work, but they can also omit qualifiers or repeat a model's error. Keep a traceable link back to the source and make it possible to rebuild or invalidate projections after correction.
Treat forgetting as a policy
Time-based decay can lower the priority of information that is less likely to be current. It should not silently rewrite history or erase records that an audit or correction flow still needs. Prefer explicit expiry, supersession, and deletion semantics that match the data's purpose.
A recent statement is not automatically more accurate than an older one. Evaluate whether it is authoritative, whether it actually supersedes the earlier statement, and what time period each statement describes.
Measure the context saved
Smaller retrieved context can reduce tokens, but the reduction alone is not success. Measure whether the answer remains correct and supported, how often required evidence is missed, whether stale facts are returned, and the latency and cost of memory construction and retrieval. Use the same questions, answer model, prompt, and token budget for each comparison.
Squish is an open-source local-first memory runtime with SQLite-backed local storage. Paid plans provide hosted memory API and MCP access; automatic cross-device sync, backups, managed AI, and team features are not included. Check the pricing page for current checkout availability and evaluate benchmarks only with their recorded protocol.