The top community-built memory and context management tools for Hermes Agent, ranked by GitHub stars. Memory is what makes Hermes self-improving — these tools extend its built-in recall with long-term semantic memory, graph-based retrieval, and cross-session persistence.
For the architecture behind these tools, read The Hermes Agent Memory Guidebook.
This universal memory layer provides multi-level storage for user, session, and agent states using temporal reasoning and multi-signal retrieval. It is available as an Apache-2.0 self-hosted server, an OSS SDK, or a managed cloud service.
OpenViking organizes context as a virtual filesystem using viking:// URIs and a three-tier loading system to reduce token consumption. This AGPL-3.0 context database offers observable retrieval trajectories and demonstrated significant task-completion gains in LoCoMo10 benchmarks.
This opinionated agent brain utilizes graph-based retrieval by extracting typed entity edges without requiring LLM calls. It features gap analysis to identify missing knowledge and enforces verified zero-leak access controls for multi-user data scoping.
Supermemory combines RAG and memory through a custom vector graph engine that supports automated fact extraction and byte-level deduplication. It features connectors for platforms like Google Drive and GitHub, sustaining sub-300ms recall at scales exceeding 100B tokens per month.
Hindsight implements long-term retain, recall, and reflect workflows to create agent memory that learns over time. The system integrates with over 25 LLM providers and is deployable via Docker, Kubernetes, or native Python and Node.js clients.
Honcho is a FastAPI-based library that enables reasoning-first memory by extracting conclusions from conversation history. It supports managed, local, and self-hosted Docker Compose deployments through Python and TypeScript SDKs.
This Elastic License 2.0 provider manages context trees using git-like versioning, including commits, branches, and merges. It features a five-tier retrieval system that achieves sub-100ms speeds, primarily through non-LLM tiers, and includes an interactive TUI.
Mnemosyne is a zero-dependency memory system that uses a single SQLite database for local storage and sub-millisecond performance. It achieves sub-linear storage growth through episodic compression and supports integrations with MCP and OpenWebUI.
This seven-layer memory system utilizes local infrastructure including Qdrant, Redis, and SQLite to manage structured facts and vector search. It employs a Ground Truth hierarchy to ensure agents prioritize specific injected context and maintains an auto-curated wiki.
Autocontext is a recursive harness that generates filesystem-first artifacts such as traces, playbooks, and context bundles. It manages context through outcome-gated promotion and ablation-backed attribution to assist agents with complex tasks.
This plugin provides lossless context management using a SQLite-backed store and a hierarchical DAG of depth-aware summary nodes. It includes specialized agent tools for searching and expanding compacted material to ensure no messages are lost.
This project enables private cross-session memory for Hermes Agent by providing a self-hosted backend for the Honcho library. It includes automated setup scripts for Docker, PostgreSQL, and Redis, supporting any OpenAI-compatible or local inference provider.
This local-first identity and memory layer provides portable state across Hermes Agent and other MCP clients. It achieves a 97.6% LongMemEval answer accuracy and supports data imports from Obsidian, Discord, and GitHub.
Plur offers a shared memory layer that stores data locally as human-readable YAML engrams. It utilizes hybrid retrieval via BM25 and local embedders to integrate with MCP tools like Claude Code and OpenClaw.
This on-device context engine operates entirely locally without API keys or cloud dependencies. It employs hybrid retrieval, including vector search and BM25, to provide memory through MCP and Claude Code hooks.
This project optimizes context for Hermes Agent and OpenClaw by using block reordering and deduplication to enable prefix reuse. These techniques reduce prefill latency by 1.5–3× and can save up to 36% of tokens during inference.
Sibyl-Memory is a file-based long-term memory system that uses SQLite with FTS5 search instead of vector databases or embeddings. It features a five-package plugin family that includes a dedicated Hermes adapter and an MCP server.
This self-maintaining plugin provides explainable retrieval by attaching a 'why_retrieved' tag to every recalled result. It includes a self-directing loop to resolve knowledge gaps and supports owner-scoping for multi-user environments.
Brainstack acts as a memory kernel for Hermes by integrating Hindsight, Graphiti, and MemPalace into a single owner. It manages state and retrieval using SQLite, Kuzu, and Chroma to deliver purpose-built evidence packets for agent queries.
This anticipatory memory system prefetches context by monitoring agent session logs and storing data locally via SQLite-vec. It is an MCP-native server designed for automated configuration with major coding agent clients.
AgentCairn uses an Obsidian vault as a source of truth, storing long-term memory as inspectable Markdown files. It functions without a daemon or opaque databases, utilizing DuckDB solely as a replaceable retrieval cache.
This layer resolves live workspace state into verified facts before the context window opens, resulting in up to 94% fewer prompt tokens. It uses SQLite and FTS5 for local-first persistence and supports MCP-native integration via stdio or SSE.
Keep provides reflective memory using semantic search, BM25 full-text ranking, and graph traversal. It features automated indexing for multimedia and codebases, using edge tags to create bidirectional relationships between notes.
This tool reduces prompt token overhead for Hermes Agent v0.14.0 by using local BM25 ranking for tool-schema selection. It includes a dashboard to track estimated token savings and supports native tool search.
This local-only plugin utilizes a columnar .lbdb embedded graph database to manage agent recall. It features a 1-10 importance ranking system to surface high-priority memories during context prefetch alongside BM25 keyword search and optional GLiNER2 entity extraction.
A Git-backed context layer that stores identities, projects, and workflows in reviewable Markdown files for coding agents. It provides a deterministic kernel for hash-checked proposals and receipts, supporting Hermes alongside Claude Code and Codex.
This episodic memory plugin implements a pull-model with an on-demand summarizer and support for real, immediate event deletion. It includes a trace.jsonl audit log to provide a transparent record of every prefetch and memory operation.
XMemo provides durable cross-session memory for Hermes through pre-turn context injection and session-end snapshots. It includes a local SQLite reliability layer and a write outbox to manage tools for searching, remembering, and updating agent state.
This SDK and MCP server enables Hermes agents to handle payments using USDC and stablecoins across EVM, Solana, and eight other chain families. It allows APIs to charge for access and agents to pay for services using only a few lines of code.
A reusable plugin designed for local capability indexing and routing of skills, scripts, runbooks, and MCP configurations. It offers five native tools for searching and fetching information, with support for opt-in automatic generation of tool routing notes.
This plugin integrates Exabase M-1 to provide self-organizing long-term memory for Hermes agents. It includes configurable query expansion and result reranking to optimize the retrieval of manually stored or searched memories.
An AI-native memory provider that defaults to AES-256-GCM encryption at rest for stored data. It supports HNSW ANN, sparse, and FM-index retrieval modes and can be executed via native Rust FFI or as an HTTP daemon.
This plugin provides private local code memory for repositories, enabling semantic search and grounded Q&A through a local SourceVault server. It adds deterministic slash commands and LLM-callable tools for models that support structured tool use.
Penfield offers 17 tools for memory management and knowledge graph traversal using a hybrid search of BM25, vector, and graph methods. It also features automatic context checkpointing to preserve information during message compression.
A file-based curation loop that manages memory hygiene by identifying rot and contradictions in versioned Markdown files. It uses five slash commands to capture and recall data, proposing evidence-cited diffs for user review without requiring a database.