YantrikDB Memory for Hermes
Give Hermes Agent persistent, maintainable memory with traceable recall, duplicate cleanup, and contradiction handling.
- Source repo
- yantrikos/yantrikdb-hermes-plugin
- Stars
- ★ 97
- Last updated
- 16d ago
- License
- MIT
- Primary language
- Python
- FA score
- 72/100 · Some gaps
At a glance
- How it runs
- Works with
- Portable with changes
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Hermes users who want an agent to retain preferences across sessions and retrieve them later.
- Not a fit if
- Teams seeking a direct plugin for Claude Code, Cursor, or another non-Hermes assistant
- Teams needing cross-process shared memory without deploying the HTTP backend
- Website
- yantrikdb.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-04latest release v0.28.0 (2026-11-03)
- Source review
- 72/100 · Some gaps
What does this agent do, and when should you use it?
This standalone plugin connects YantrikDB memory to NousResearch Hermes Agent through Hermes's memory-provider interface and lifecycle hooks. Its default backend runs the engine in process, so it does not require a separate server; an HTTP backend can instead connect to yantrikdb-server. The plugin exposes tools for storing, recalling, relating, resolving, and maintaining memories, with hooks for events such as session end and pre-compression. Recall results include why_retrieved scoring reasons, while embedded mode also supports correction chains and historical queries. Installation requires the plugin and YantrikDB engine to be importable from Hermes's Python environment; this is not a direct plugin for other agent platforms.
The plugin routes Hermes memory operations through YantrikDB: yantrikdb_remember writes memories, yantrikdb_recall retrieves them with why_retrieved reasons, yantrikdb_relate creates links, and yantrikdb_conflicts and yantrikdb_resolve_conflict expose and resolve contradictions. yantrikdb_think runs maintenance. Lifecycle hooks can inject high-salience memories before compression, mirror writes to Hermes's built-in memory, and trigger maintenance at session end. When enabled, additional tools cover skills, tasks, recent turns, knowledge gaps, triggers, and knowledge packs. Embedded mode supports correction writes with supersedes_rid, chain inspection with include_superseded, historical recall with as_of, and edge inspection through yantrikdb_links; the HTTP backend does not support these relationship features.
- Hermes users who want an agent to retain preferences across sessions and retrieve them later.
- Long-running Hermes coding agents that need high-salience constraints injected before context compression.
- Teams maintaining user or project facts who need to surface duplicates and contradictions.
- Users who need to audit corrections by asking what was current at a past time; this requires the embedded backend.
- Operators of multiple Hermes agents who want shared server-backed memory or engine reuse among providers in one host process.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install yantrikdb-hermes-pluginpip install yantrikdb
- Environment variables
YANTRIKDB_URLYANTRIKDB_TOKEN- Config files
~/.hermes/hermes-agent/venv/bin/activate~/.hermes/.env
How do you install or deploy this agent?
For the package install, install into the Python environment Hermes uses, register the provider, and select it in Hermes:
pip install yantrikdb-hermes-plugin
yantrikdb-hermes install
hermes memory setupTo install the plugin source through Hermes instead, activate Hermes's environment and install the engine separately:
source ~/.hermes/hermes-agent/venv/bin/activate
hermes plugins install yantrikos/yantrikdb-hermes-plugin
pip install yantrikdb
hermes memory setupThe HTTP backend requires a separately deployed yantrikdb-server, a token, and YANTRIKDB_MODE=http, YANTRIKDB_URL, and YANTRIKDB_TOKEN settings in ~/.hermes/.env. The default embedded backend needs no server, token, or GPU.
How do you use this agent?
After registration, select the provider and check its availability:
hermes memory setup
hermes memory statusThen ask Hermes to remember or recall information in a conversation; the plugin handles the operation through tools such as yantrikdb_remember and yantrikdb_recall. To isolate memories by owner, enable scoping in ~/.hermes/.env:
echo "YANTRIKDB_OWNER_SCOPING=true" >> ~/.hermes/.envSkill tools are disabled by default and can be enabled explicitly:
echo "YANTRIKDB_SKILLS_ENABLED=true" >> ~/.hermes/.envWhat are this agent's strengths and limitations?
- The default embedded backend runs without a separate server, token, or GPU; an HTTP backend is available for multi-instance deployments.
- Recall results include
why_retrievedreasons that show why memories ranked. - Includes contradiction handling, deduplication maintenance, and newer correction-chain and historical-recall features.
- The README documents runnable recall benchmarks, test coverage, and a reproducible provider-comparison probe.
- Its native integration is for Hermes Agent; other assistants require a different integration, and the README's MCP option is a separate YantrikDB project.
- The plugin and engine must be installed in Hermes's actual Python environment, so setup and environment troubleshooting involve several steps.
- Correction chains, historical queries, and link traversal are embedded-only; they are unavailable through the HTTP backend.
- In the README's 1,000-record comparison probe, YantrikDB stored 256 of 1,000 attempted writes before an engine queue-full error; that probe used an older version combination.
How does this agent compare with similar options?
Compared with Hermes's built-in file memory, this plugin adds YantrikDB semantic retrieval, scoring explanations, contradiction handling, and maintenance. The README also reports a reproducible probe against ByteRover, Hindsight, Holographic, Honcho, Mem0, OpenViking, RetainDB, and Supermemory; some providers could not be verified without credentials or endpoints, and measured results are limited by the probe corpus and query format.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| YantrikDB Memory for Hermes This agent | 72 · Some gaps | Agent plugin / skillFree | ★ 97 | 16d ago | Python | — |
| SwarmVault | 75 · Good | CLIFree + model costs | ★ 737 | 4mo ago | TypeScript | Codex · Claude Code |
| Obsidian Wiki — A Digital Brain You Grow with Your AI Agent | 73 · Some gaps | Agent plugin / skillFree | ★ 3.8k | today | Python | Codex · Claude Code |
| Due Diligence Agents | 69 · Some gaps | CLIFree + model costs | ★ 139 | 3d ago | Python | Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
The installer refuses to overwrite an existing target unless --force is supplied; the README gives uninstall steps and describes the HTTP token setup, embedded default, and dependency ranges. Deductions: the source sample does not show encryption, retention, or sensitive-data policies for stored memories; forced replacement and uninstall delete the plugin directory without a recovery mechanism, though the README warns about replacement; publisher identity is unverified, and maintenance attribution is mainly in project files and author metadata.
CI is configured for linting, type checks, and tests across multiple Python versions, with a separate real-engine contract job; version-sync checks and dependency upper bounds address drift. Deductions: this static review cannot verify that the jobs pass; the sample contains extensive historical compatibility details without enough implementation to confirm end-to-end consistency. The CLI gives specific messages for an existing target, an invalid Hermes path, and UI runtime errors, but the sample does not cover every runtime failure.
The documentation clearly targets Hermes users and covers embedded single-host use, HTTP for multiple instances, owner scoping, long-running deployments, and agent fleets, with configuration examples. Deductions: boundaries and trigger conditions are spread across versioned README sections, and compatibility with a particular Hermes version still requires user verification; the sample does not include the full tool and automation implementations.
The repository includes a README, CLI, project configuration, CI and publishing workflows, an MIT license, and a changelog reference; it provides concrete install, update, uninstall, environment, optional embedder, and limitation guidance. Deductions: the README is long and accumulates details across releases, so key options require searching across sections; the author, issue path, and release process indicate responsibility, but no response commitment or support timeline is stated.
The product integrates memory writes, recall, contradiction tracking, maintenance, and explainable recall into Hermes, returning information such as why_retrieved; the README gives context-cost estimates, benchmark references, and usage examples. Deductions: benefits and competitor comparisons are largely project-authored claims, and the sample cannot independently establish practical gains for ordinary users; some cost and performance figures are explicitly not newly measured.
README guidance on installation, versions, and dependencies is corroborated by the CLI, pyproject, and CI workflows; the repository also identifies benchmark tooling and version-sync checks. Deductions: many performance, scale, and competitor claims depend on reports, tests, or external material absent from this sample; the README sometimes distinguishes measured results from older estimates, but evidence trails are not equally clear for every claim.
- Memories may contain sensitive user information; the provided source sample does not show encryption, retention periods, or data-erasure policy. Verify how database files and backups are protected before deployment.
- Performance, scale, and competitor comparisons are mainly project-authored claims, and some tool-cost figures are explicitly not newly measured. This static review did not execute code or verify test results.
- Updating with --force replaces the plugin directory; the README advises backing up custom files stored there.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v0.28.0 v0.28.0 — correction chains, historical recall and link traversal
Release notes excerpt
Memory can now be corrected, not just appended to. - yantrikdb_remember(supersedes_rid=…) links a correction to what it replaces; default recall returns only the current head. include_superseded / as_of give the full chain or the state at an instant; new yantrikdb_links tool traverses edges. Embedded backend only; the HTTP backend refuses loudly (#98) - model2vec is a manifest hard dependency, so it survives Hermes env regeneration (#92, #96) - sentence-transformers extra gated off android/emscripten so it no longer breaks Hermes' workspace resolve (#99, #100) …View on GitHub ↗ -
v0.27.1 v0.27.1 — a missing embedder dependency says which interpreter looked where
Release notes excerpt
A missing embedder dependency now names the interpreter and sys.path it searched (#92), and the README documents that embedder dependencies installed by hand are dropped when Hermes regenerates the plugin environment. Making model2vec a manifest dependency is still open in #92. No functional change otherwise.View on GitHub ↗ -
v0.27.0 v0.27.0 — a multiplexed gateway keeps each profile's YantrikDB settings to itself
Release notes excerpt
Pin unchanged (yantrikdb>=0.12.1,!=0.15.0,!=0.15.1,!=0.15.2,<0.24.0). Gated against engine 0.23.1. **Anyone running a multiplexed Hermes gateway with more than one profile should take this release.** Validated live on Hermes 0.19.0: before it, a secondary profile's memories went into the launch profile's store, even when that profile's own .env configured a different one. Install note. Hermes' memory loader prefers plugins/memory/yantrikdb inside the Hermes package over $HERMES_HOME/plugins/yantrikdb. If you ever used the deprecated …View on GitHub ↗
FAQ
Does the default installation require a separate database server?
Do I need an API key or GPU?
Can memories be isolated by user?
YANTRIKDB_OWNER_SCOPING=true. It is off by default; under the default settings, the base namespace remains included in recall.Does the HTTP backend support historical correction queries?
Is YantrikDB Memory for Hermes free?
Is YantrikDB Memory for Hermes open source?
Is YantrikDB Memory for Hermes safe to use?
Is YantrikDB Memory for Hermes still maintained?
Who makes YantrikDB Memory for Hermes?
What is YantrikDB Memory for Hermes's official website?
Thanks — we use this to improve the page.