Releases·Tool Release
Open-source memory service Athena tiers agent memory with topic-chained, Ebbinghaus-curve forgetting
The builders of Athena, an open-source agent memory service, argue that history retention alone doesn't give agents memory — memory must be earned via a promotion mechanism. Athena organizes memories into three tiers and uses topic-shifted chains scored on an Ebbinghaus forgetting curve so only frequently used, important information persists long-term.
1 engagement1 sourceThu, Sep 24, 2026, 12:08 PM
AI Overview
- Athena, an open-source agent memory service, uses three tiers: recent-event window, summarised topic chains, and a long-term graph of recurring people, projects, and tools.
- Chains are cut by topic shift via cosine similarity between consecutive turns, not volume or TTL.
- Closed chains carry a heat score, Heat = I_base * exp(-ΔT / (τ * S)), with τ defaulting to 24 hours and S rising on each retrieval so use flattens forgetting.
- Only chains hot enough on the forgetting curve are extracted into the long-term graph, making forgetting the default and remembering earned.
- In a demo, one correction stopped an agent opening a mechanic's briefs with already-known car details, shifting them to lead with undone work.
Related
Social·Community
My agent's memory after 25 weeks: four graphs, no LLM in the promotion path, and three ways it broke
A developer details the agent memory system built over 25 weeks, replacing summarize-and-stuff-context with tiered memory (working, episodic, semantic) and nightly consolidation. The design deliberately excludes LLMs from the promotion path, relying on recurrence and access-pattern signals, and splits memory into four graphs. An audit surfaced three promotion defects plus a generalized failure mode where Hebbian reinforcement forms attractor basins that break retrieval.
A developer details the agent memory system built over 25 weeks, replacing summarize-and-stuff-context with tiered memory (working, episodic, semantic) and nightly consolidation. The design deliberately excludes LLMs from the promotion path, relying on recurrence and access-pattern signals, and splits memory into four graphs. An audit surfaced three promotion defects plus a generalized failure mode where Hebbian reinforcement forms attractor basins that break retrieval.
12 engagement1 sourceredditSun, Sep 13, 2026, 02:52 AM
Products·launched
Knowl: Agent memory that retires stale facts when they change
Knowl is an agent memory system that not only stores facts but also updates or removes them when they become stale, preventing AI agents from giving conflicting answers. It solves the problem of long-term context management for AI agents, ensuring they always use the latest information. Aimed at developers using AI agents in production.
Knowl is an agent memory system that not only stores facts but also updates or removes them when they become stale, preventing AI agents from giving conflicting answers. It solves the problem of long-term context management for AI agents, ensuring they always use the latest information. Aimed at developers using AI agents in production.
1 engagement1 sourcehackernewsSat, Aug 22, 2026, 02:13 PM
Developments·Paper
What Should an Agent Forget? Separating What Is Stored from What Is Used
An arxiv paper introduces RD-Forget, a training-free framework for persistent language agents that separates what an agent stores from what it uses when answering. A retained source archive preserves observations, while a query-conditioned memory view controls which stored facts influence a given answer. This addresses the problem that superseded facts can mislead current-state answers yet remain essential for historical queries.
An arxiv paper introduces RD-Forget, a training-free framework for persistent language agents that separates what an agent stores from what it uses when answering. A retained source archive preserves observations, while a query-conditioned memory view controls which stored facts influence a given answer. This addresses the problem that superseded facts can mislead current-state answers yet remain essential for historical queries.
0 engagement1 sourcearxivWed, Sep 9, 2026, 02:48 PM
Products·building
Collapse Aware AI: Memory middleware that gates agent behavior
Collapse Aware AI, Evolution 2 is middleware that sits between an AI agent and its memory, ensuring that retrieved history must 'earn' the right to influence the agent's actions. It uses a pipeline of retained history, bounded retrieval, candidate behaviors, governance, and final selection to filter memory impact. This solves the problem of irrelevant or harmful memory affecting agent decisions, improving reliability for developers building AI agents.
Collapse Aware AI, Evolution 2 is middleware that sits between an AI agent and its memory, ensuring that retrieved history must 'earn' the right to influence the agent's actions. It uses a pipeline of retained history, bounded retrieval, candidate behaviors, governance, and final selection to filter memory impact. This solves the problem of irrelevant or harmful memory affecting agent decisions, improving reliability for developers building AI agents.
4 engagement1 sourceredditWed, Aug 12, 2026, 10:52 PM
Social·Community
Developer wrestles with persistent agent memory design
A developer building persistent memory into an AI agent found that storing full conversation history made the agent worse, pulling outdated context and contradicting itself. They are now researching proper memory design, including MemGPT.
A developer building persistent memory into an AI agent found that storing full conversation history made the agent worse, pulling outdated context and contradicting itself. They are now researching proper memory design, including MemGPT.
46 engagement2 sourcesredditMon, Aug 24, 2026, 03:41 AM