By Michiel Horstman · Founder, 4M Labs · Updated Oct 20, 2026
A context window is the material available to a model during a request. Persistent memory is information retained outside that request and retrieved when a later task needs it. A larger context window can hold more material at once, but it does not by itself decide what should persist, how it should be corrected, or who may retrieve it.
Use context for active work
The model needs the current task, relevant instructions, and supporting evidence in its active context. Sending an entire history may be expensive and distract from the question; sending too little may omit a necessary detail. The right amount depends on the model, task, and prompt.
Use memory for selected carryover
Persistent memory can make prior decisions or preferences available to a later session. It needs a lifecycle: source tracking, retrieval, correction, expiration, and deletion. Summaries can help manage volume, but should remain connected to the evidence they represent.
Test the trade-off
Compare a no-memory approach with a memory-backed one on the same tasks. Measure answer quality, source support, missing evidence, stale-memory errors, token use, latency, and cost. Do not infer a token-saving percentage from context size alone; memory construction, retrieval, and additional agent turns have costs too.
Squish is an open-source, local-first memory runtime whose local database uses SQLite. Host integration and network activity depend on release and configuration. Paid plans provide hosted memory API and MCP access; automatic cross-device sync, backups, managed AI, and team features are not included. Check the current documentation for setup and the pricing page for checkout availability.