An agent may convert one successful run into a reusable rule that was never valid outside that exact context. WikiTrace stores the compact lesson alongside a longer account and a link to the original execution trace, allowing the agent to inspect source evidence before reusing its memory. Across eight controlled benchmark suites, the paper reports mean…
Inspect the ClaimsCase · DiggingBeagle record
WikiTrace reduced agent-memory regressions by preserving links to original execution traces
An agent may convert one successful run into a reusable rule that was never valid outside that exact context. WikiTrace stores the compact lesson alongside a longer account and a link to the original execution trace, allowing the agent to inspect source evidence before reusing its memory.
Scope
Agent-memory experimental evaluation under selected benchmark tasks; no adversarial poisoning or secure retention trial.
Assessment profile
Assessment method- Setting
- evaluation
- Exploitation
- not established
- Evidence
- Grade B
- Remediation
- not applicable
- AI role
- AGAINST AI
Basis and provenance
Memory-linked execution traces improve selected task outcomes, with no attacker poisoning or confidential trace-retention audit. An impact band is not assigned because the available Claims do not establish a comparable verified severity or quantified harm.
Assessed Nov 12, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R239-011-CORE · WikiTrace links reusable agent memories to complete original trajectories and lowers memory-induced regressions
At a glance
Source links help correct context drift but are not a substitute for authenticating the source of an agent memory.
Read the implicationsA compromised trace, false success label or credential retained within memory may defeat a seemingly evidence-linked result.
Limits and uncertaintyFull account
An agent may convert one successful run into a reusable rule that was never valid outside that exact context. WikiTrace stores the compact lesson alongside a longer account and a link to the original execution trace, allowing the agent to inspect source evidence before reusing its memory.
Across eight controlled benchmark suites, the paper reports mean task score 90.2 with WikiTrace versus 87.0 for the same insight-only memory and 79.5 without memory. The result measures recall and task utility, not security against forged experiences: original traces can themselves be poisoned, mislabelled or contain secrets. Linking to an execution record improves auditability only if the record is trustworthy and access-controlled.
Source links help correct context drift but are not a substitute for authenticating the source of an agent memory.
A trace link improves inspectability, but an attacker-controlled log can still contain poisoned instructions or private data. Unlike pre-admission falsification approaches, provenance alone does not reject a false success label. The reported task gains are benchmark utility measures, not a demonstration of poisoning resistance or safe enterprise retention.
Mechanism and trust boundary
- 01
An agent compresses a successful trajectory into a reusable memory insight.
- 02
WikiTrace retains a link to the longer original execution instead of trusting only the abstracted lesson.
- 03
Later tasks consult that provenance to correct context-specific rules, contingent on the original trace being valid.
Claims & evidence
CLM-R239-011-COREWikiTrace reports 90.2 average task score across eight suites, compared with 87.0 using insights alone and 79.5 without memory in its selected evaluation.supported
Basis: reported finding
- supportsWikiTrace links reusable agent memories to complete original trajectories and lowers memory-induced regressionsprimary disclosure
Section 4.2 Table 1, gain/regression comparison and trace-descent analysis
Implications
Source links help correct context drift but are not a substitute for authenticating the source of an agent memory.
Controls and mitigations
- Sign and restrict access to source traces, validate original outcomes before promoting a memory and test explicitly poisoned histories.
Unknowns and contradictions
- A compromised trace, false success label or credential retained within memory may defeat a seemingly evidence-linked result.
Sources and citation
Material revision history
- Nov 12, 2026 · Published version · first publication · revision 241