Context, retrieval, and memory

Use the interactive context budget explorer to compare selection policies and inspect omissions before running the fixtures.

State is what the runtime needs to continue. Context is what the model receives on one request. Memory is information selected for retention across requests or sessions. Retrieval chooses evidence; it does not make that evidence authoritative.

python -m examples.harness.context
python -m pytest tests/test_learning_context.py -q

The example stores synthetic facts in SQLite under tenant and session keys, reconstructs the store, retrieves evidence, and builds a bounded context with source IDs. Tests reopen it in another interpreter and attempt cross-session/tenant retrieval. The demo prints only Alice's fact, with source manual-1; Bob's fact must not appear.

Choose a context policy

Policy	Useful when	Tradeoff and verification
Full transcript	Short tasks need exact history	Cost grows; inspect the wire, including schemas
Recent window	Old details are dispensable	Can lose approvals/constraints; retain obligations explicitly
Summary	History is large and semantic compression helps	Can alter facts; evaluate with real models and original evidence
Selective retrieval	Relevant evidence can be identified	Retrieval can miss facts; track source IDs and absence
Persistent memory	Future work needs selected facts	Needs isolation, expiry, correction, and consent/authority rules

This example budgets characters, not tokens, and selects standalone facts, not raw tool-message history. A production transcript window must preserve valid tool request/result pairs. Required evidence is selected first; if it cannot fit, the builder fails visibly. Omitted source IDs are reported. The tests do not establish semantic summary quality or real-model retrieval accuracy.

The store assumes tenant/session values came from trusted request context. Passing model-supplied tenant IDs to it would defeat isolation. Storage persistence is not an authentication system, and the demo's SQLite lifecycle is not a distributed memory service.

Use prompt-growth evidence to understand costs and the RAG arena for fixed-corpus task mechanics. Extend this lesson with an early required fact, irrelevant recent evidence, a corrected source, and two users with identical search terms. Track what is omitted and why.

Return to the build path.