FrontierMemory · evidence-audited research lineage

Last checked October 25, 2026. Live diagram · Official FrontierMemory post · Letta research index.

Edge semantics: solid = FrontierMemory explicitly names or links source; dashed = thematic chronology from Letta’s own research index (not a blog citation); dotted = a recall-oriented benchmark the FrontierMemory post explicitly contrasts. Exception: MemGPT → Sleep-time is thematic within the organization, not claimed as a FrontierMemory citation.

flowchart TB
    subgraph ORIGIN["Letta earlier research · thematic unless cited"]
        memgpt["MemGPT (2023)"]
        sleep["Sleep-time Compute (2025)"]
        filesystem["Letta Filesystem (2025)"]
        contextbench["Context-Bench (2025)"]
        skilluse["Skill Use / Context-Bench (2025)"]
        skilllearning["Skill Learning (2025)"]
        token["Token-space learning (2025)"]
        contextrepo["Context Repositories (2026)"]
        memorymodels["Memory Models (2026)"]
        trajectory["Trajectory format (2026)"]
    end
    subgraph CODING["Task benchmark sources"]
        deepswe["DeepSWE (2026)"]
        skillsbench["SkillsBench (2026)"]
        automation["AutomationBench (2026)"]
        apex["APEX-Agents (2026)"]
        enterprise["EnterpriseOps-Gym (2026)"]
    end
    subgraph EVALUATION["Experience and methodology"]
        harborindex["HarborIndex (2026)"]
        sweinteract["SWE-Interact (2026)"]
        studying["Machine Studying (2026)"]
    end
    subgraph OTHER["Recall benchmark contrasts"]
        locomo["LoCoMo (2024)"]
        longmemeval["LongMemEval (2024)"]
    end
    subgraph GOVERNOR["FrontierMemory 1.0"]
        frontier["FrontierMemory 1.0 (2026)"]
    end
    sleep --> frontier
    filesystem --> frontier
    deepswe --> frontier
    skillsbench --> frontier
    automation --> frontier
    apex --> frontier
    enterprise --> frontier
    harborindex --> frontier
    sweinteract --> frontier
    studying --> frontier
    locomo -. contrasted .-> frontier
    longmemeval -. contrasted .-> frontier
    memgpt -. related .-> frontier
    contextbench -. related .-> frontier
    skilluse -. related .-> frontier
    skilllearning -. related .-> frontier
    token -. related .-> frontier
    contextrepo -. related .-> frontier
    memorymodels -. related .-> frontier
    trajectory -. related .-> frontier
    memgpt -. related .-> sleep
    classDef target fill:#fff5eb,stroke:#bc773e,stroke-width:3px
    class frontier target

Primary sources and basis

Caveats

• FrontierMemory v1.0 is a published research blog/repository, not independently confirmed peer-reviewed proceedings paper.
• The source blog has 15 adapted tasks; it reports 219 relevant among 665 prior trajectories. The v1.0 released corpus on our cluster had 674. Letta’s blog does not reconcile the different counts.
• Direct source benchmarks: DeepSWE, SkillsBench, AutomationBench, APEX-Agents, EnterpriseOps-Gym. Historical trajectories: HarborIndex and DeepSWE; simulated users: precedent SWE-Interact.
• Gold memory for each task is Letta’s reported validation design, not evidence that our own earlier 7B runs were capable of using it.
• An explicit link does not prove an implementation was copied. Organizational background links are not treated as reference-list citations.
• Original broad graph: historical visualization; archived previous Mermaid.

Publication venue mapping

The page’s venue section maps the 19-entry original Markdown list plus the MSR 2027 Data & Tool Showcase.

• FORGE Data & Benchmarking: November 15, 2026, benchmarking papers.
• EASE 2027: January 15 abstract / January 22 paper for research.
• AGENTVERIFY dedicated CFP: November 13, conflicts with an older ICSE aggregate calendar; reconfirm before submission.
• ICSE Industry Challenge: Agent Memory in the Wild is one of its actual listed challenges, but a challenge-specific solution is required.
• AgenticSE workshop deadline from old MD remains unconfirmed (site requires JavaScript); ICLR workshop contribution dates are workshop-specific.