What is Holographic?
Holographic is a memory provider for Nous Research's Hermes Agent. It keeps facts in one local SQLite file and recalls them before each turn using FTS5 keyword search, entity names, trust scores and HRR (holographic reduced representations, a way of binding words into fixed-size vectors with simple algebra). By default it makes no network or model calls.
It shipped inside Hermes core until Hermes v0.21.7 removed the bundled copy on 11 October 2026. Nous Research handed it to a community maintainer, itpartypattaya, who publishes it as the catalogue plugin named holographic; the same author has started a standalone port, holographic-memory, for Claude Code and Codex.
The plugin is MIT-licensed, written in Python and released often: version 0.26.4 came out on 3 November 2026. Several unrelated GitHub projects use the word holographic for agent memory; this is the one Hermes installs.
Is Holographic worth it?
Holographic treats memory as a small database of explicit facts rather than a pile of embedded chat logs. The agent writes facts through a fact_store tool with nine actions, including add, update, contradict and remove; each change becomes a dated version, so the store can say what held on a past date, and forget erases a fact with its history. Nothing calls an LLM to extract facts. An optional auto_extract setting, off by default, saves user sentences that state a preference or decision in English, Russian or Chinese at a reduced trust of 0.4.
Recall is lexical first. Before each turn the message's words are matched against FTS5 across 18 languages; up to five facts that cover enough of the message, plus facts about named people, go into the prompt marked as data, not instructions. HRR vectors of 1,024 dimensions contribute 30% of the ranking. Trust scores rise by 0.05 when the agent marks a fact helpful and fall by 0.10 when unhelpful, and facts below 0.3 are not recalled. Writes that look like instructions or credentials are refused.
The numbers come from the maintainer's own test sets, not LongMemEval or LoCoMo. On 450 queries in 18 languages the target fact was returned 430 times with search by meaning off and 433 with it on. Recall without the reranker took 3.7 ms at the median and 9.2 ms at p95 on a Ryzen 7 2700; switching on the optional cross-encoder reranker raised that to 176 to 184 ms on the same CPU. Those figures say the store is fast and precise on facts it holds, but nothing about long conversations, which is what the standard benchmarks test.
The risks are organisational. The provider has one maintainer and a 112 KB README of edge cases, and schema changes have consequences: after a first merge, version 0.19.0 and later lock older releases out of the file. The Claude Code and Codex port is a week old. For Hermes users it is the most transparent option; for anyone else Mnemosyne or claude-mem have larger user bases.
Where it earns its place
A family running a shared Hermes bot in a Telegram group
Remember each person's preferences without mixing them up
Set scope to user so each member gets a separate SQLite file, and leave auto_extract on for preferences. The reranker gate cut wrong-person recalls from 28 to 6 in the maintainer's tests, at the cost of a local reranker server.
A privacy-conscious consultant using Hermes on a laptop
Keep client facts on the machine with no cloud calls
Install from the catalogue with hermes plugins install holographic and keep embeddings and rerank off. The database file is owner-readable only (0600) and nothing leaves the machine; recall takes single-digit milliseconds.
A developer splitting work between Claude Code and Codex
Share one memory of project facts between both agents
The standalone holographic-memory port adds a hook and an MCP server with 23 tools over one store in ~/.holographic. It is early 0.x software from the same maintainer, so treat it as a trial.
Before you commit
- Setup
- hermes plugins install holographic, then hermes config set memory.provider holographic. Hermes installs it automatically for profiles that already used the bundled version; existing memory_store.db files are kept.
- Integrations
- Hermes Agent memory provider API; optional OpenAI-compatible embeddings endpoint and llama-server cross-encoder reranker; a separate port for Claude Code and Codex over MCP.
- Data and privacy
- One SQLite file in the Hermes home, permissions 0600. Nothing is sent anywhere unless embeddings or rerank is switched on, in which case fact text and messages go to the endpoint you name. Each fact version records its source chat and author; forget deletes them.
- Cost at scale
- Free to run. The optional reranker adds a llama-server process and 176 to 184 ms per turn on a Ryzen 7 2700 CPU, or about 32 to 35 ms on a GTX 1050 Ti.
- Limits
- No LLM extraction, so the agent must call fact_store or facts are missed. Two-question messages are handled only by an experimental setting. Benchmarks are on the maintainer's own sets.
What’s good
- Fully offline by default; no API key or model calls
- Every fact versioned, with as-of queries and hard forget
- Refuses to store instructions or credentials
- Keyword recall in 18 languages, including Russian and Chinese
What’s not
- No LongMemEval or LoCoMo result
- One community maintainer since October 2026
- Mainly useful inside Hermes Agent
- Downgrading below 0.19.0 after a merge switches memory off
Best for
Hermes Agent users who want inspectable, offline fact memory and are willing to have the agent write facts deliberately.
Holographic alternatives
Holographic: common questions
- Is Holographic worth it?
- Holographic is the careful, offline choice for Hermes users who want a fact store they can audit, but it is a one-maintainer plug-in with no standard benchmark and little use outside Hermes. We rate it 3.6 out of 5.
- How much does Holographic cost?
- Free and open source under the MIT licence. Optional search by meaning and reranking call endpoints you choose and pay for.
- What are the downsides of Holographic?
- No LongMemEval or LoCoMo result; one community maintainer since October 2026; mainly useful inside Hermes Agent; downgrading below 0.19.0 after a merge switches memory off.
- Who is Holographic best for?
- Hermes Agent users who want inspectable, offline fact memory and are willing to have the agent write facts deliberately.
- How much does Holographic cost for a team or in production?
- Free to run. The optional reranker adds a llama-server process and 176 to 184 ms per turn on a Ryzen 7 2700 CPU, or about 32 to 35 ms on a GTX 1050 Ti.
- Is Holographic safe for company data?
- One SQLite file in the Hermes home, permissions 0600. Nothing is sent anywhere unless embeddings or rerank is switched on, in which case fact text and messages go to the endpoint you name. Each fact version records its source chat and author; forget deletes them.
- What are the best Holographic alternatives?
- The closest alternatives we have reviewed are Mnemosyne, claude-mem and Mem0.
Answers checked Nov 6, 2026.