Fly Labs
Ask FlyBot

AI memory with biological decay (52% recall)

Most RAG setups fail because they treat memory like a static filing cabinet. When every transient bug fix or abandoned rule is stored forever, the context window eventually chokes on noise, spiking token costs and degrading the agent's reasoning.This implementation experiments with a biological approach by using the Ebbinghaus forgetting curve to manage context as a living substrate. Memories are assigned a "strength" score where each recall reinforces the data and flattens its decay curve (spaced repetition), while unused data eventually hits a threshold and is pruned.To solve the "logical neighbor" problem where semantic search misses relevant but non-similar nodes, a graph layer is layered over the vector store. Benchmarked against the LoCoMo dataset, this reached 52% Recall@5, nearly double the accuracy of stateless vector stores, while cutting token waste by roughly 84%.Built as a local first MCP server using DuckDB, the hypothesis is that for agents handling long-running projects, "what to forget" is just as critical as "what to remember." I'd be interested to hear if others are exploring non-linear decay or similar biological constraints for context management.GitHub: https://github.com/sachitrafa/cognitive-ai-memory

Hacker NewsToolAISource
0Sign in to voteCopy link

FL score

72

out of 100

Verdict

VALIDATE

medium confidence

Competition

No competitor data yet

Trend

No signal yet

A memory management layer for AI agents that uses decay curves and graph topology to prune noise and reduce token costs in long-running systems.

The pain

RAG systems accumulate garbage, deprecated rules, transient fixes, abandoned context, that bloats token usage and degrades reasoning quality. A 10-month project with 50k stored facts becomes a liability, not an asset. Teams either manually curate (expensive) or accept degradation (silent failure).

The gap

Existing vector stores treat memory as static. Semantic search alone misses logical neighbors. No standard approach exists for *forgetting* in production AI systems. The builder has a working prototype (52% recall, 84% token reduction on LoCoMo), but it's unclear if this generalizes beyond benchmark datasets or if users will adopt a new memory abstraction.

Build angle

Start by embedding this as a drop-in MCP server for Claude/other agents running multi-week projects. Target teams already paying for high token usage (consulting firms, research orgs, code generation shops). Measure actual token savings and reasoning quality on *their* real workflows, not benchmarks. If 3, 5 customers see 60%+ token reduction and measurable quality gains, the unit economics unlock. The biological metaphor is elegant but secondary, the real win is operational cost and reliability.

Strengths

  • Solves a real, measurable problem: token bloat in long-running agents. 84% reduction is concrete.
  • Prototype is built and benchmarked. Not vaporware.
  • Graph layer addresses a genuine gap in vector-only retrieval (logical neighbors).
  • Local-first, DuckDB-based architecture avoids vendor lock-in and appeals to privacy-conscious teams.
  • Ebbinghaus framing is memorable and differentiates from 'just prune old stuff.'

Risks

  • 52% recall on LoCoMo is good for a benchmark, but LoCoMo is synthetic. Real-world recall may be lower; users may not tolerate forgetting critical context.
  • Adoption friction: requires teams to trust a new memory layer. Existing RAG workflows are entrenched. Why switch if current system 'works'?
  • Decay curve tuning is domain-specific. A rule that's 'transient' in one project is 'foundational' in another. One-size-fits-all decay will fail.
  • No clear pricing or business model. Is this a library, a service, or a consulting play? Unclear path to revenue.
  • Competitive risk: major RAG vendors (Pinecone, Weaviate, LlamaIndex) could add decay/pruning features in weeks. Defensibility is low.
  • The 84% token reduction claim needs validation on production workloads, not just benchmarks. Real-world savings may be 20, 40%.

Questions about this idea?

FlyBot reads the scoring and gives you a second opinion on “AI memory with biological decay (52% recall)”.

Open FlyBot