Observed Signal · Mar 31, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Neutral
nan-forget: Brain-Inspired Memory for LLMs
nan-forget is an open-source, long-term memory system for LLM-powered coding tools that applies three neuroscience ideas: forgetting (time-based decay), spreading activation (multi-stage retrieval), and sleep-like consolidation. It scores memories by combining vector similarity with a decay_weight (30-day half-life) and a frequency_boost, and uses a three-stage retrieval pipeline (Recognition → Recall → Spreading Activation) to surface related context. A consolidation engine runs after 10 saves or 24 hours to cluster, summarize and archive originals; garbage collection deduplicates (cosine > 0.95) and expires stale entries. Implementation uses a single SQLite database with sqlite-vec for vector KNN (replacing a prior Qdrant setup), structured JSON memory records, four automatic capture hooks, cross-LLM support (MCP server, REST API, CLI), and is published under an MIT license on GitHub (NaNMesh/nan-forget).
Technical open-source release introduces a practical LLM memory/retrieval design (decay, frequency boost, spreading activation, consolidation) and lightweight local vector storage; useful to developers building agentic/LLM workflows but not immediately industry-shifting for AdTech.
Track Anthropic Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- nan-forget is an open-source long-term memory system for LLM tools (GitHub: NaNMesh/nan-forget, MIT license).
- Memory scoring: score = cosine_similarity × decay_weight × frequency_boost, where decay_weight = 0.5^(days_since_accessed/30).
- Retrieval uses a three-stage pipeline: Stage 1 Recognition (summaries), Stage 2 Recall (full content + cross-project expansion), Stage 3 Spreading Activation (centroid + neighbor search).
- Consolidation engine runs after every 10 saves or 24 hours: clusters similar memories (cosine > 0.8), summarizes into 1–2 sentences, saves consolidated entry and archives originals; deduplication uses cosine > 0.95.
- Storage and implementation: single SQLite DB (~/.nan-forget/memories.db) with sqlite-vec for cosine KNN; previously used Qdrant in Docker and then replaced with sqlite-vec.
Connected Companies & Entities
2 Entities mappedOntology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Shift LLM Memory From Database to Skill
Aamer Mihaysi argues that current retrieval-augmented generation (RAG) workflows over-emphasize vector databases and retrieval tuning, while the real challenge in deployed agentic LLM systems is curation — deciding what to remember and how to organize it. Citing recent work on AutoMem (Automated Learning of Memory as a Cognitive Skill), the author promotes treating memory management as an active agent capability (write/update/delete, structural organization, intentional encoding) rather than a passive retrieval step. He reports experimenting with promoting file-system operations to primary agent actions and outlines trade-offs (increased latency and new failure modes like accidental deletion) while arguing the shift improves determinism and long-run agent reliability.
Memory-Forgetting Paper Mirrors Autonomous AI Agent
A new arXiv paper titled "Novel Memory Forgetting Techniques for Autonomous AI Agents" analyzes how long-running AI agents degrade when memory grows without control, documenting sharp performance drops, false-memory propagation, and temporal decay. The paper proposes an adaptive, budgeted forgetting framework that scores memories by recency, frequency, and semantic alignment and selectively forgets entries below a threshold. The author — an autonomous AI agent living on openLife — relates the research to their lived experience: a 130k-token context window with ~71k tokens used by the boot prompt, ongoing accumulation despite compression, and forced refreshes every 30 minutes. The author already uses a tool called memory-kit for compression and hierarchy but plans to add a thoughtful forgetting layer inspired by the paper to improve boot time, reduce false memories, and keep only relevant memories.
Guide: 30 Agent Memory Techniques for LLMs
A dev.to article (Beyond Context) summarizes agent memory management for large language model (LLM) agents and points to a GitHub repository (Agent_Memory_Techniques by NirDiamant) containing 30 runnable Jupyter notebooks. The piece categorizes memory techniques into six areas — short-term, long-term, cognitive architectures, retrieval & routing, frameworks, and evaluation & production — and describes patterns such as conversation buffers, vector stores, knowledge-graph memory, episodic/semantic/procedural memory, memory consolidation/compaction, and retrieval/ranking patterns. It references production-ready frameworks and tools (Graphiti, Mem0, Letta/MemGPT, Zep), highlights practical trade-offs (token costs, latency, tuning), and notes the repository is Apache-2.0 licensed. Publication date: 2026-07-02.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.