Pick the safest write policy for an agent that stores extracted memories on every turn

MCQ·Hard·4.0 · 0·~1 min·Asked atAnthropic OpenAI Perplexity

Attempt it

**Two extremes fail in opposite directions.** Writing everything unconditionally causes memory bloat: duplicates pile up, contradictions accumulate, retrieval gets noisier turn over turn. Writing nothing (or only what the agent volunteers) causes amnesia: useful facts the agent should have remembered are lost. The robust default is a four-step pipeline. - **Extract.** Run a small extraction model over the user message and (optionally) the agent response. Produce candidate facts in a structured shape: `{subject, predicate, object, timestamp, confidence}`. - **Deduplicate.** Check each candidate against existing memories via embedding similarity and an LLM-judge near-neighbor check. If the candidate is already in the store, do not write a duplicate; optionally update the timestamp on the existing entry. - **Resolve contradictions.** If the candidate contradicts an existing memory, decide which wins. Usually the more recent fact, but with explicit rules for safety-critical or user-confirmed facts. - **Threshold-gate.** Write only candidates whose extraction confidence is above a threshold (typically 0.7-0.8). Below threshold, drop without writing. This is the default policy in Mem0 and Zep. Storing raw transcripts (option 2) drowns retrieval in chatter; tool-gated writes with a default of nothing (option 3) is what Letta does but it requires strong system-prompt guidance to avoid amnesia.

**The right write policy is a four-stage pipeline, not a single decision.** The three rejected options each correspond to a specific known failure mode in production memory systems. **Why append-only fails.** Memory stores under append-only writes grow linearly with conversation length. Within months of daily use, a typical user has 5-10K facts and many are duplicates of each other in slightly different phrasings. Retrieval starts pulling near-duplicates that crowd out unique signal, and contradictions accumulate silently, the store says both 'user lives in Portland' and 'user lives in Seattle' because they moved and no one reconciled. Deduplicating only at retrieval time does not fix the contradictions, only the duplicates, and at the wrong stage. **Why storing full transcripts fails.** Memory is supposed to be a distilled, retrievable representation, not a verbatim log. Full transcripts drown retrieval in conversational chatter ('thanks', 'sounds good', 'one second'). The signal to noise ratio is wrong. Real agent products that tried this in 2023-2024 universally moved to extracted-fact memory by 2025. **Why tool gated with default nothing fails by default.** When the agent has to actively choose to remember something, it under-remembers. Useful background facts the user mentions in passing get lost. Letta uses this style and is rigorous about it, but it requires careful system-prompt guidance to function and is more fragile than extraction-side automation. **The robust pipeline:** - Extract candidate facts via a small dedicated model. - Deduplicate via embedding similarity + LLM-judge near-neighbor check. - Resolve contradictions explicitly: most recent wins for ordinary facts, user confirmed wins for safety, with an audit trail of supersessions. - Threshold-gate on extraction confidence to drop noisy candidates. Mem0 and Zep both ship this as their default. The robustness is in the pipeline structure, not in any single step.

Quick reference

TL;DR

The safe write policy has four steps: extract candidate facts, deduplicate against existing memories, resolve contradictions explicitly, then write only above a confidence threshold; this is what Mem0 and Zep

Memory aid

Easy to grasp

Imagine taking notes during a long conversation. If you scribble down every single thing the other person says, your notebook fills up with junk and you cannot find the important parts later. If you wait and decide later whether to write anything, you forget what mattered. The middle ground is what good journalists do: listen for the real points, check whether you already have that point in another note, reconcile if the new one disagrees with the old one, and write the new note only if you are confident it is true and useful. That same discipline is the safe write policy for an agent's memory.

Key concepts

Context engineering

Detailed answer & concept explanation

~6 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Quick scan · 2–4 min

Everything important, quickly.

3 min: map each rejected option to its failure mode; name the four stages of the robust pipeline; explain extraction model specialization, dedup versus contradiction distinction, and per-type confidence thresholds; reference Mem0, Zep, Letta.

Real-world examples

Real products, models, and research that use this idea.

Mem0 in 2026 ships the four-stage extract dedup resolve threshold pipeline as default with tunable confidence thresholds per memory type.
Zep implements the same pipeline backed by a temporal knowledge graph where contradictions become superseded edges with explicit timestamps.
Letta uses tool-gated writes (core_memory_append, archival_insert) paired with strong system-prompt guidance about when the model should remember something.
Anthropic's Claude memory features in 2026 expose memory writes as tool calls but recommend an extraction layer on the application side.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you tune the confidence threshold for the threshold-gate stage?

QHow does the contradiction-resolution rule change for safety facts versus preferences versus transient state?

QWhat is the architectural difference between extraction-driven and tool-driven write policies, and when does each fit?

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Appending every extracted fact unconditionally. The store grows fast, contradictions accumulate, retrieval gets noisy, and the agent's behavior degrades over months even though no single write looked wrong.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

References

Primary sources. Skim if you want the original framing.