Useful memory without lost obligations

This worked application uses explicit planning assumptions to demonstrate FIELD. Quantities below are inputs, calculated implications or proposed targets; they are not reported customer results. The engagement decision remains conditional on the evidence described.

Planning basis and engagement boundary

F — Frame the outcome

Maintain useful continuity across months while respecting changed preferences, obsolete sources and deletion requests.

Value measure: Task completion with correct current context, correction burden and stale-memory incidents.

I — Investigate the workflow

Classify durable facts, preferences, transient context and operational records. Observe which forgotten or stale details actually harm the user's task.

First evidence task: Delete a fact and query through fresh, cached and summarized context paths.

E — Engineer the solution

Use scoped/versioned memory and context selection; keep authoritative operations outside model summaries and budget retrieval/context separately.

The implementation assessment below makes the authority, workload, resource and failure assumptions concrete. Its numbers are calculations from the stated scenario, not a vendor performance guarantee.

Contract

Define memory categories: explicit preferences, authoritative business facts, inferred hypotheses, conversation history and durable task obligations. Establish who may write/read each category, expiry, correction and deletion behavior. Do not merge all categories into one vector store without semantics.

Implementation plan

Store typed records with provenance, subject/tenant, effective/observation time, confidence and supersession links. At each turn, retrieve relevant authorized records, refresh volatile facts and combine them with current task state and recent conversation. Compaction creates a derived summary; approvals/action outcomes remain in a separate authoritative workflow store.

Sizing

With 200 sessions×2,000 tokens, raw history totals 400,000 tokens before the new request. An illustrative context plan might select 2,000 history/memory tokens, 3,000 evidence tokens and 1,000 instructions/task tokens plus output reserve. Those numbers are a budget, not a guarantee that retrieval preserves all required information. Context length affects prefill and KV memory; persistent index/storage grows independently.

Conflict and deletion

A new explicit user preference supersedes an old inferred preference. A current authoritative business record overrides an old conversational assertion. Deletion propagates to summaries, vectors and caches with a recorded lifecycle; task audit retention may have distinct requirements that must be established. Do not allow late replay to recreate deleted memory.

Compaction failure

Keep a machine-readable obligations list and exact action references so an omitted exception cannot authorize a wrong action. Tests intentionally remove important facts from the summary and check that the system refreshes or asks rather than guessing.

Evaluation

Compare no-memory, raw-history and retrieved-memory baselines on held-out longitudinal tasks, including changed preferences, similar users, conflicts, stale facts and deletion. Measure useful retention, stale-memory errors, cross-user leakage, cost and latency. A memorized expected answer is leakage, not useful memory.

L — Launch with evidence

Test contradictory updates, deletion propagation and cross-user isolation, then compare task usefulness and correction effort against a limited-memory baseline.

Critical decision condition: Deleted or unauthorized memory must not be resurrected through a cache or summary.

Attach raw artifacts, the tested configuration, population, denominator, date and limitations to the release record. Missing or stale evidence holds the dependent decision. A favorable value forecast cannot offset a failed critical condition.

D — Develop ownership and reuse

A product owner defines memory semantics and users retain appropriate control; operators own deletion/reconciliation across stores.

Candidate reusable asset: Memory provenance and deletion test contract.

Commercial decision and next commitment

Accountable decision owner: Assistant product owner with data owner.

The initial commitment is the bounded journey above, with the stated exclusions. Confirm access, customer staffing and acceptance ownership during discovery; price and schedule unresolved dependencies explicitly. A new source, tenant class, language, action or operating envelope requires an impact review rather than silent scope growth.

Approve only the next exposure level supported by the evidence. The next useful investment is: delete a fact and query through fresh, cached and summarized context paths.

Method and references

Apply the FIELD method, economic model and evidence rules. Methodological and technical influences are listed in the source register.