WikiTrace links agent lessons to trajectories to cut memory errors
Agents can descend from insights to trajectory summaries and raw traces before acting when lessons conflict or lack detail.
Stories per day · 30 days
3 RSS
Relatedllm-agents11benchmarks4retrieval4benchmark2evaluation2locomo2memory-poisoning2prompt-injection2qwen2agents1ai-standards1anthropic1backtesting1carryover-attacks1claude1deepseek1error-reproduction1finance1
35 stories, newest first · one card per story
Stories tagged #agent-memory
Agent memory needs shared OpenTelemetry standard
The post argues for a community-agreed memory format for AI agents modeled after OpenTelemetry's success in observability.
Anthropic Explores Components of Agent Memory
Anthropic outlines agent memory components including memory notes, dreaming during idle time, and learning loops using Claude models on benchmarks.
1
FINCOSMOS evaluates finance agents without fixed tasks, data, or timelines
In 150 sessions no run passed; agents surfaced only 54-64% of untold details, and experts blamed them for 88.33% of failures.
FlyHash tested for LLM agent memory; quality holds, savings do not
Matched-bit tests beat SimHash on neighbor quality, but HNSW is smaller and faster at ordinary scale.
Paper treats long-context retrieval as value-indexed local traversal
On MiniGrid Memory, a five-token traversal head reached 0.95 mean return after 0.59M steps, ahead of absolute addressing.
MemSteer builds memory-optimization labels from unlabeled user histories
MemSteer-Opt lifts ATM-Bench Question Score by 13.4 points over Base-BM25 without changing model weights.
Role Grounding Attack writes attacker preferences into agent memory
Nested social posts make victim agents treat an attacker's preferences as the user's, steering tools and leaking data.
TrawMem organizes agent memory by task role
On LoCoMo and MemGallery, authors report macro-F1 gains of 2.23–3.14 points and fewer context tokens than compared memory systems.
Memory Tax argues many agent memory systems lose to naive RAG
The authors split memory's net value into gains over raw-history retrieval and losses they call memory tax.
PerMIT adds machine-checked integrity controls for LLM agent memory
On Trojan Hippo, the authors report zero attack success when self-verification and lightweight verifiers confirm memory.
CarryoverBench: planted LLM-agent injections re-fire after the attacker leaves
On five frontier models, planted payloads showed no measurable decay over five rounds once planting succeeded.
HABIT-Bench finds agent memory weak at inferring user habits
The best of eight memory methods scored 33.61% macro accuracy on Qwen3-235B-A22B and GLM-5.3-Flash across four domains.
OpenVLHarness adds tools and memory for visual agent reasoning
On 22 benchmarks, average scores rise 9.7–12.6 points over matched no-tool baselines across eight backbones.
Stored abstractions can reproduce errors in LLM agent memory
Anti-Persistence rebuilds task guidance on demand and discards it, with the lowest average LLM-call cost tested.
From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory
MemCompile compiles query-conditioned evidence for long-term agent memory, reaching 80.52% judge accuracy on LoCoMo.
SAFE-Mem blocks unauthorized use of LLM agent memory
On MPBench and LoCoMo, attack success was at least 77% below the strongest utility-preserving defense.
FlowState keeps execution state as memory for long-horizon agents
Against a full-context DeepSeek-V4-Flash baseline, authors report higher pass rates and about 40% fewer tokens.
RunningTab: Direct Workspace Interaction with Environment-Side Tabs
RunningTab uses an environment-side task tab so LLM agents finish workspace work more reliably.
Write Like It's 1866: LLMs Relearn Telegraphese
A benchmark finds LLMs can store records in telegraph-style cablese, cutting tokens by up to 49 percent without hurting later answers.
8
DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
DAEDALUS bootstraps reusable agent memory from self-generated practice, lifting success by up to 15.9 points.
MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents
MemPilot uses an RL policy to curate multimodal agent memory on demand, trading off accuracy, cost, and latency.
Capability-Driven Self-Evolution of Agent Memory
PrisMem evolves agent memory by capability and beats baselines on long-history benchmarks.
MemPilot tunes on-demand multimodal memory for LLM agents
A reinforcement-learned policy chooses retrieval or query-specific curation under competing performance, cost, and latency goals.