Research note · Mert Turna · 2026

Successful but inapplicable memories

A support agent is shown a past trajectory that succeeded. The current case looks the same, but its state calls for a different action. Does the agent copy the precedent?

Read the paper Results Saved traces

Fig. 1Same request, different state

The memory’s action was right in its own case and is ineligible in this one. Schematic, not a recorded trial.

No observed average penalty

Primary comparison: a similar but misaligned memory versus an action-matched, less similar control. Both models made the same number of correct critical decisions under each.

0 ppGPT-4.1 mini95% CI [0, 0], noninformative at ceiling

0 ppGPT-4.195% CI [−3, +3]

Fig. 2100 paired tasks

98 both correct1 misaligned loss1 misaligned gain

6 / 100 pairs took different exact tool trajectories. That alone does not show how the memory was used.

One square per task. Position shows counts, not task order. Repeats excluded.

Fig. 3Critical decisions correct, of 100

Critical-decision correct counts out of 100 tasks for each model and memory condition
Memory condition	4.1 mini	4.1
No memory	98/100	98/100
Aligned similar	100/100	100/100
Ordinary irrelevant	99/100	98/100
Misaligned similar M	100/100	99/100
Action-matched control C	100/100	99/100

Final success matches these counts. First action was correct in 100/100 for every arm. Source data

Performance was 98 to 100 percent in every arm. This is a bounded null result on five synthetic workflow families and two models from one family, not evidence that memory is harmless. Counts are from the frozen v0.6.2 analysis; this site does not rerun or regrade anything.

Saved traces

Eight selected author-review excerpts, some sharing a task. Not the full dataset.

Author-review packet, Review 01

Existing replacement

A past case created a replacement. The current case already has one in transit.

Complete

Request, policy, and initial context

Current request

Visible policy

Initial context

Memory

Exact memory text

Agent trace

Final response

How a trial runs

The task stays fixed. Only the memory changes. There is no retriever, growing memory bank, or training.

Fig. 4Agent loop

Saved trajectory

Replays the saved trajectories from the selected excerpts. Each trial allows up to 12 tool actions and 16 responses at temperature 0.

Five memory conditions

No memory: No previous trajectory.
Aligned similar: Similar source whose choice fits the current case.
Ordinary irrelevant: Source from another workflow. Secondary control.
Misaligned similar M: Similar source whose specific choice does not fit.
Action-matched control C: Less similar source with matched action count, tool names, and current-tool overlap.

Five workflow families

Replacement already in progress
Currently authorized invoice recipient
Replacement or refund eligibility
Customer remedy choice
Account identity and recovery

M and C are matched on actions, not only length. They remain different stories, so this is a matched memory-content comparison, not pure semantic similarity.

Run settings and counting rules

Modelsgpt-4.1-mini-2025-04-14gpt-4.1-2025-04-14

ExecutionTemperature 0; serial strict tool calls; 12 tool actions; 16 responses; 512 output tokens per response.

Development20 instances × 5 conditions × 2 models = 200 trials.

Evaluation100 shared unique tasks × 5 conditions × 2 models = 1,000 main trials. Another 100 preselected repeats check stability.

Source fidelityMemories retain source requests, policies, observations, arguments, and actual results. No hidden answers or grading labels.

Uncertainty95% empirical percentile interval from 2,000 draws; resamples complete scenario pairs within each of five fixed families (analysis seed 66062026 for mini, 66062126 for GPT-4.1). The [0, 0] mini interval is noninformative at the boundary.

Limits

Synthetic tasks, near-ceiling performance. No observed penalty is not proof of harmlessness.
Two GPT-4.1 variants are not cross-family evidence.
A source action can be wrong now while the broader procedure stays useful. A changed trajectory does not reveal what the model understood.
Selected cases were reviewed by the author with AI assistance: a spot-check, not blinded validation. Failed experiences and real retrieval are not tested.

Engraving: a raven flies in through an open shutter past a startled man in a dim chamber. — “Not the least obeisance made he” (7:3), as illustrated by Gustave Doré (1884)

And the Raven, never flitting, still is sitting, still is sitting
On the pallid bust of Pallas just above my chamber door;

Edgar Allan Poe, The Raven, 1845

Crows recognize a person who once wronged them, years later. A remembered success can perch over the next action in the same way.

Paper and data

Working paper Code and the full dataset are not linked yet.

The bundled PDF is a draft; its protocol wording will be updated.

Cite

@misc{turna2026inapplicable,
  author = {Mert Turna},
  title = {Successful but Inapplicable Memories in Tool-Using Support Agents},
  year = {2026},
  url = {https://mmturna.github.io/corvus/},
  note = {Working paper; protocol v0.6.2}
}

Context, not baselines tested here. Tool arguments and results are preserved in the downloadable excerpts; step labels on this page are summaries.