MEMORY AS A RISK CONTROL
Which mistakes does memory change?
Compare mistake probabilities and costs on the same trips, with memory and review on or off.
Fresh retrieval experiment · database, context & filtered memory →
What belongs in context, memory and the database?
| Layer | Contents | Failure to measure |
|---|---|---|
| Authoritative database | Current identities, prices, permissions, ledgers and policy versions | Wrong entity, stale revision, duplicate action |
| Working context | Current task, trusted constraints and the few records needed for this decision | Truncated evidence, omitted constraint, attention overload |
| Retrieved long-term memory | Source-linked procedures and prior lessons, scoped to the task and time | Irrelevant retrieval, obsolete guidance, scope leakage, conflicting authority |
A database can store both current records and long-term memory. Here, retrieved procedures supplement authoritative records. The experiment measures the supplied guidance package, including its added context.
Placement and retrieval experiments
Hold the model, tasks, token allowance, corpus and critic fixed. Compare direct database access, the same guidance pinned in context, unfiltered semantic retrieval, and retrieval filtered by entity, authority and valid time. Randomize condition order and keep discovery cases separate from held-out evaluation.
Primary outcomes: each policy-violation probability, loss above a declared dollar threshold and mean loss. Record both corrected mistakes and newly introduced mistakes. Measure end-to-end latency, retrieval latency, tokens and cost as secondary tradeoffs. The fresh three-placement experiment is available through the link above. Unfiltered semantic retrieval remains a proposed additional arm; the archived results on this page compare the earlier memory package.
Test stale facts, superseded rules, missing retrievals, look-alike identities, irrelevant context, poisoned instructions and critic overreach. Add new failure hypotheses from development traces, then test them on fresh held-out episodes.
Combining real traces, stress tests and unknown failures
For each scenario s, estimate a separate loss-exceedance probability from real traces and from simulations. Use declared traffic weights w, real-trace share ρ and unseen-failure mass ε:
P(L > t) = (1 − ε) Σₛ wₛ [(1 − ρ) pₛᵢₘ,ₛ(t) + ρ pₒᵦₛ,ₛ(t)] + ε q(t)
Each tested scenario gets a Beta–Binomial update. The untested component q(t) spans 0 to 1 unless validated evidence narrows it. Vary ε to see its effect on the tail-risk range. Source and traffic weights specify how test results transfer to the workload.
Explore source-mixture assumptions →Adjust unseen-failure mass interactively →Protocols and primary documentation
Risk-focused memory study design ↗
SurrealDB scopes and retrieval traces ↗ · Mem0 filtered search ↗ · Letta memory architecture ↗