What types of memory errors do chatbots actually make?

Chen and colleagues build a taxonomy of "memory misalignment" — cases where a chatbot's stored memory diverges from what the user intended or expects — from a two-phase, IRB-approved study: a memory database of 457 entries from 28 participants, plus a three-week diary study collecting 304 reports from 32 participants. The resulting 14 misalignment types (labeled I1–I14) are organized across three lifecycle stages adapted from Jones et al.'s model of agent memory: intake (input filtering and extraction), storage and management (consolidation and organization), and retrieval and interpretation (recalling and applying stored memories). The excerpt details the first four intake-stage types. I1, over-inference, is "an error where the system elevates low-frequency, anomalous, or transient utility-driven user actions into stable, longitudinal preferences" — the example given is a chatbot logging one-off searches ("AirPods pairing, removing chilli oil stains, AC power consumption") as persistent "daily areas of interest." I2, attribution confusion, is committing third-party or fictional details — from a creative-writing exercise, in the example given — to the user's own profile. I3, identity misinference, is wrongly deducing a demographic trait, such as inferring a user's gender as female when "the user issued a direct correction stating they are male." I4, hallucination, is inventing a memory with no basis in the dialogue at all, including the case of an internal system instruction ("the system must firmly maintain the current year as 2026") being logged as if it were a user trait.

Each type is cross-mapped to a category in Schacter's cognitive taxonomy of human memory errors — sins of omission (transience, absent-mindedness, blocking) and sins of commission (misattribution, suggestibility, bias, persistence). Over-inference is labeled a commission-bias error; attribution confusion and identity misinference are commission-misattribution; hallucination is commission-suggestibility. The paper's framing move is to treat a chatbot's memory failures as structurally analogous to documented human memory distortions, rather than as one-off engineering bugs, which is what licenses a unified 14-type taxonomy instead of a list of ad hoc complaints.

This gives a system-side vocabulary for a phenomenon Does ChatGPT's memory feature feel like a relationship or a machine? documents from the user's side: an expectancy violation on discovering what ChatGPT retained is, in this taxonomy's terms, likely an instance of over-inference or identity misinference surfacing to the user after the fact. It also supplies a mechanism for why What mental models do users hold about LLM memory? finds users holding incompatible models of what memory does: a user who assumes memory is session-only has no reason to expect a one-off query to become a daily area of interest, so over-inference reads as a violation rather than a feature. The retrieval-and-interpretation stage this taxonomy names also runs alongside, rather than duplicating, the gap in Why do LLM agents remember preferences but not act on them?: that paper's failure is knowing a preference but not applying it, while this taxonomy's retrieval-stage errors (not detailed in the excerpt beyond their stage label) concern applying a memory inappropriately — outdated or over-applied rather than simply ignored.

The excerpt breaks off at I4, so the remaining ten types — covering storage/management errors like temporal staleness and granularity mismatch, and the retrieval-and-interpretation errors — are named as categories but not described here, and no rate is given for how many of the 457 or 304 logged items fell into each type. The taxonomy also doesn't itself specify which failures are more frequent or more damaging; that weighting, and the design responses to it, are addressed separately in the paper's co-design and speed-dating work. The implication the taxonomy does support: memory misalignment is not one problem (retrieval accuracy) but at least three, split by where in the memory lifecycle they originate, so a fix aimed at one stage — better retrieval ranking, say — will not touch intake-stage errors like over-inference or hallucination.