Long-term memory for your AI coding agents. Tell your agent something once — a decision, how your setup works, a correction — and it remembers it in every future session, in any project, with any of your agents.
- Remembers across sessions. Decisions and their reasons, your environment, your preferences, people and projects, corrections. Your agent searches the memory before answering and saves what lasts, so you stop repeating yourself.
- One memory for all your agents and models. Claude Code, Codex and opencode can share the same memory: what one saves, the others find. Tested with models from Anthropic, OpenAI, Google and Meta.
- Saves tokens. Instead of pasting your notes or re-explaining context every session, the agent retrieves only what the question needs — typically 200–900 tokens, whether the memory holds 50 entries or 50,000, because search walks a tree instead of reading everything.Honest numbers , including where it doesnot save.
- Understands time. When something changes ("I moved to Valencia", "staging is on PostgreSQL 17
now"), the old memory is kept as history, flagged
OUTDATED, and searches answer with the current one first. Where you live, where you work and your stack are tracked automatically; any other correction is linked when the agent saves it withreplaces. When a question is ambiguous ("Luis" — which one?), it asks instead of guessing. - Portable. Your memory is one folder of plain text files. Copy it to another computer, back it up, or put it under version control — it works wherever you put it.
- Local, private and readable. No cloud, no database, no account: IHMT stores everything on your disk and sends nothing anywhere. (The memories your agent retrieves reach its model like any other context.) Every memory is a text file you can open, and each person who installs IHMT starts with their own, empty memory.
Compatibility. Officially supported: Claude Code. Also tested: Codex (CLI and the
ChatGPT desktop app — setup) and opencode
(setup). IHMT is a standard stdio MCP server, so any
agent that supports local MCP servers should work — GitHub Copilot, Antigravity, Cursor, Windsurf,
Gemini CLI, Claude Desktop… — and INSTALL.md knows how to configure them, but we have
not tested those yet. All of them can share one memory.
| You need | Why |
|---|---|
| Python 3.10 or newer | IHMT is written in Python |
| git | to download IHMT and keep it updated |
| An AI agent that can run commands (Claude Code, Codex, opencode, Copilot in agent mode…) | it installs IHMT and then uses the memory |
| Internet, during the installation | to download the code and the MCP package; not needed afterwards |
Missing Python or git? Your AI agent installs them for you (it is instructed to ask you first). Python goes in your user folder, with no administrator password, so nothing system-wide changes. On a brand-new Mac, git may need one click: Apple shows a window asking to install its command-line tools.
On a Mac, note that the python3 that comes with macOS is version 3.9, which is too old — that is why
your agent may say Python is missing even though python3 exists.
You do not need administrator rights, a database, an account, or any paid service beyond your agent. Tested on macOS and Linux (Ubuntu); on Windows the instructions are included but not tested yet. Details: GUIDE.md §3.
Paste this into the AI agent you want to give a memory to — Claude Code, Codex and opencode are tested; GitHub Copilot, Antigravity, Cursor, Windsurf, Gemini CLI, Claude Desktop and other MCP clients should work too:
Install the IHMT memory MCP server for me from https://github.com/gonzaroman/IHMT-MEMORY — follow the instructions in its INSTALL.md.
The agent follows INSTALL.md: it downloads IHMT to ~/IHMT-MEMORY, keeps your
memory in ~/.ihmt, registers the server with itself only, adds the usage instructions, and
tells you what it did. Then start a new session so the memory tools load.
Using several agents? Paste the same prompt in each one, whenever you want. If IHMT is already installed — say you have used it with Claude for months and now want it in Codex — the agent finds that installation, updates it if it safely can, and connects to the same memory, so it knows what you told the others from day one.
It needs an agent that can run terminal commands or edit files; chat-only assistants in a browser cannot install anything.
Manual install #
Requirements: Python 3.10+, git, and your agent's CLI.
git clone https://github.com/gonzaroman/IHMT-MEMORY.git ~/IHMT-MEMORY
cd ~/IHMT-MEMORY
python3 -m venv .venv
.venv/bin/pip install -r requirements-mcp.txt
mkdir -p ~/.ihmt
Claude Code
claude mcp add ihmt-memory -s user -e IHMT_HOME="$HOME/.ihmt" -- "$PWD/.venv/bin/python" "$PWD/mcp_server.py"
claude mcp list # ihmt-memory … ✔ Connected
The server name must come before -e. Then append
templates/memory-instructions.md to ~/.claude/CLAUDE.md.
Codex — codex mcp add ihmt-memory --env IHMT_HOME="$HOME/.ihmt" -- "$PWD/.venv/bin/python" "$PWD/mcp_server.py",
then add default_tools_approval_mode = "approve" to the [mcp_servers.ihmt-memory] table in
~/.codex/config.toml (above its env table) and append the template to ~/.codex/AGENTS.md.
Details.
opencode — add an "ihmt-memory" entry ("type": "local", "command": [<python>, <mcp_server.py>],
"environment": {"IHMT_HOME": <memory folder>}) to the "mcp" object of
~/.config/opencode/opencode.json, and append the template to ~/.config/opencode/AGENTS.md.
Details.
Windows, the project scope, the graphical setup and troubleshooting are all in the guide.
New here? Read GUIDE.md — everyday use, step by step, with real outputs. The rest
of this README is the technical reference.
A universal, domain-agnostic long-term memory for LLMs, stored as a recursive tree of plain files on the local disk. No vector database, no server, no third-party dependencies — Python 3.10+ and the standard library.
Instead of embedding everything into one flat index and scanning it, IHMT organizes knowledge into a
tree: raw text leaves at the bottom, recursive JSON summaries above them, and a single root.json
trunk at the top. A query walks that tree — root → branch → branch → leaf — so the number of files
opened grows with the depth of the tree (≈ beam × log_B(n)), not with the amount stored.
| Flat RAG | IHMT | |
|---|---|---|
| Retrieval cost | scan / ANN over all n chunks | beam × log_B(n) file reads |
| Structure | none — a bag of vectors | explicit hierarchy, inspectable |
| Chunking | fixed character windows | syntax-, scene- and date-aware |
| Stale facts | served silently | superseded, dated, and flagged |
| Ambiguity | returns a plausible guess | asks you for a clue |
| Storage | binary index | UTF-8 .txt + JSON you can read |
python3 gui.py # graphical interface: set up, browse, inspect
python3 init_ihmt.py # or from the terminal: create ./ihmt_memory
python3 main.py demo # full walkthrough in ./demo_workspace
python3 -m unittest discover -v # stdlib only; MCP tests skip without the SDK
Then use it on your own material:
python3 main.py ingest ~/notes ~/project/src/Main.java
python3 main.py consolidate --force
python3 main.py search "how did we handle stock reservations"
python3 main.py ask "Luis" # interactive clue loop
python3 main.py conflicts # what changed over time
As a library:
from ihmt import IHMT
memory = IHMT.initialize("./workspace")
memory.ingest_file("examples/InventoryService.java")
memory.ingest_file("examples/journal_personal.txt")
memory.flush() # close the tree up to the root
answer = memory.search("reserveStock soft hold")
print(answer.best.content) # the leaf
print(answer.best.path) # ['root', 'N1-software.java-…', 'L-software.java-…']
print(answer.node_reads) # how many branch files were opened
for notice in memory.notices():
print(notice) # "On 2024-03-11 you said user location = 'Madrid', but on 2026-02-03 you updated to 'Valencia'."
python3 gui.py # opens a browser at 127.0.0.1
python3 gui.py --path ~/my-memory --port 8765 --no-browser
Still zero dependencies — the server is http.server from the standard library, it listens only on
the loopback interface, and every /api/* call needs the random token carried in the URL it opens.
Four screens: Set up (pick the memory folder with the system dialog, create the store, choose between per-project and global registration, preview the exact command or JSON before anything is written), Explore (collapsible tree down to the stored text, with supersession notices), Diagnose (a search that reports confidence, files opened vs. total, and the descent path), and Timeline (active vs. historical values and the detected contradictions). The interface is bilingual (ES/EN) and read-only over the memory: it never deletes or edits a leaf.
mcp_server.py exposes the tree to Claude Code as nine tools in three families: long-term memory,
project indexes and session scratch memory. The core stays dependency-free; the
SDK is an optional extra:
python3 -m venv .venv
.venv/bin/pip install -r requirements-mcp.txt # mcp[cli]>=2.0
For a single project, copy .mcp.json.example to that project's .mcp.json and fill in the absolute
paths; Claude Code asks you to approve it on the next session there (claude mcp list shows it as
Pending approval until then):
{
"mcpServers": {
"ihmt-memory": {
"command": "/absolute/path/to/IHMT-MEMORY/.venv/bin/python",
"args": ["/absolute/path/to/IHMT-MEMORY/mcp_server.py"],
"env": { "IHMT_HOME": "/absolute/path/to/IHMT-MEMORY" }
}
}
}
or in one command:
claude mcp add ihmt-memory --scope user \
-e IHMT_HOME=/absolute/path/to/IHMT-MEMORY \
-- /absolute/path/to/IHMT-MEMORY/.venv/bin/python /absolute/path/to/IHMT-MEMORY/mcp_server.py
IHMT_HOME selects the store ($IHMT_HOME/ihmt_memory), created on first use. Point every project at
one shared directory for a single cross-project memory, or give each project its own.
| Tool | Behaviour |
|---|---|
search_memory(query, clue=None, detail="compact") |
Walks the tree. Compact output: the best memory with its date and OUTDATED notices, one line per other match;detail="full" adds ids, tree paths and excerpts. An ambiguous query returns anAMBIGUOUS block listing the candidates instead of guessing — call again withclue . A query that matches nothing says so, and so does one whose closest entry shares only a stray word with it (NOT FOUND ). |
save_memory(content, domain="general", content_type="auto", replaces="") |
Classifies, splits and stores the text, extracts dated facts, and keeps the tree consolidated. Reports how it was filed and — if the save contradicts something remembered earlier — the notice to relay to the user. With replaces (a few words describing an earlier memory) the save is recorded as its correction: the old memory is flaggedOUTDATED and searches answer with the new one first. |
mark_outdated(old_id, new_id) |
Flags one memory as corrected by another, when save_memory found several candidates forreplaces and listed their ids. |
project_map(path, detail="files", subpath="") |
Compact map of a codebase: files with their size in tokens and, with detail="symbols" , each method with its line range. Built from the sync manifest, without opening leaves. |
find_code(query, path, scope="main", subpath="", clue=None, max_tokens=1500) |
Returns only the symbol that answers the query, as file:first-last + code.scope ismain (skip tests),test orall . |
read_file(path, force=False) |
Reads a file and remembers what it handed out this session: a repeated read answers UNCHANGED or only a unified diff. |
note(text) /recall(query, clue=None) |
Session scratch memory: survives a context compaction, disappears when the session ends. |
digest_output(text, label="output") |
Condenses a long log to its first lines, errors, failures, test totals and last lines; the full text stays recallable. |
Long-term memory saves tokens between sessions. The project tools save them within one, where the
cost is reading the same files again and again. ProjectIndex (ihmt/project_index.py) keeps a
private store per project under $IHMT_PROJECTS_DIR (default $IHMT_HOME/ihmt_projects):
- one leaf per symbol —
code_chunk_mode="symbol", so a lookup returns a method, not a file; - checksum sync on every call — size and mtime first, SHA-256 only for what moved; changed files are re-ingested, removed ones deleted, and the branches rebuilt. Code is never served stale;
- path-ordered branches — each file is stamped with its rank in path order, so every branch covers neighbouring files and its summary stays meaningful for the descent;
- code-aware ranking — the navigator filters by
scope/path_prefixfrom the catalog, prefers the file a query names, demotes tests unless asked and demotes import lines.
Measured on a 55-file Spring Boot project (8 typical questions): reading the files that hold the
answers costs 3,613 tokens; find_code returns the exact method for all 8 in 1,071. The map of the
project costs 660 tokens against 13,157 to read it whole.
Those savings are against an agent that reads whole files. In an A/B test with 22 real headless
Claude Code sessions, Claude preferred batched grep/ sed -n and never called the project tools on
its own; forcing them made sessions 42–71 % more expensive. Keeping the server enabled costs about
460 tokens per conversation, since Claude Code loads MCP tools on demand. IHMT's main value is memory
between sessions; treat the project tools as optional, and do not mandate them in CLAUDE.md.
The server transparently supports MCP SDK 2.x (MCPServer), 1.x ( FastMCP) and the standalone
fastmcp package.
The usage instructions in your ~/.claude/CLAUDE.md (template) tell Claude Code when to reach for each tool: search before answering anything that
depends on earlier sessions, save durable facts with their date, never guess on AMBIGUOUS, always
relay OUTDATED, and never store secrets.
Everything lives in one relocatable directory:
ihmt_memory/
root.json # the trunk: domains, topics, top branches, counters
layer_0/<domain>/*.txt # the leaves: raw UTF-8 text + a strict JSON header
layers/1/*.json # branches: summaries of leaves
layers/2..N/*.json # branches: summaries of summaries
state/catalog.json # index: id -> path, domain, parent, timestamp
state/facts.json # the fact timeline
ihmt.config.json # branch factor, token budgets, backend
root.json lives inside ihmt_memory/ so the store is self-contained: copy the directory and the
memory travels with it.
A leaf is a normal text file that describes itself, so it stays meaningful even if the catalog is lost:
<<<IHMT-META
{
"leaf_id": "L-software.java-0001-84130ee123",
"timestamp": "2026-09-09T17:45:00Z",
"data_type": "CODE",
"domain": "software.java",
"tags": ["method:reserveStock", "class:InventoryService", "lang:java", "type:code"],
"parent_id": "N1-software.java-5571a7baac",
"span": {"start_line": 43, "end_line": 68},
"checksum": "sha256:…",
"extra": {"context": "public final class InventoryService {"}
}
IHMT-META>>>
public Optional<String> reserveStock(Sku sku, int quantity) {
…
A branch node embeds each child's title, excerpt and keywords. That is the detail that makes the descent cheap: a branch can be ranked without opening any of its children.
DomainDetector classifies each document by type (CODE, NARRATIVE, CLINICAL, PERSONAL,
PROCESS, GENERIC) and domain ( software.java, medicine.clinical, process.cooking,
personal, …) from its extension plus lexical signatures in English and Spanish. Both can be
overridden with --domain / --type.
The type selects the splitter, and every splitter obeys one invariant:
A logical block is never cut. If a single block exceeds max_tokens it is stored whole and
flagged oversized. Correctness of the block beats hitting the token budget.
- Code (
chunkers/code.py) — Python via the stdlibast; Java/JS/TS/C/C++/C#/Go/Rust/Kotlin/ Swift/PHP viaBraceScanner, a character-level scanner that tracks brace depth while skipping comments, string literals, char literals, template literals and preprocessor lines. Imports coalesce; each class/function is a block; an oversized class splitsper member , and the enclosing class header travels in the leaf'sextra.contextrather than being spliced into the text. Concatenating a file's leaves reproduces the filebyte for byte — asserted in the tests. - Narrative — paragraph-atomic, with
***,---,Chapter/Capítuloas hard boundaries. Only a paragraph larger thanmax_tokensis split, and then at sentence boundaries. - Temporal (journals, chats, clinical records) — one dated entry is atomic and a change of date is a hard boundary, so a leaf never mixes two encounters.The date found in the text becomes the leaf's timestamp , which is what makes recency weighting meanwhen something was true rather than when it was ingested.
- Process (recipes, protocols, runbooks) — steps and ingredient lists stay attached to their heading.
Leaf ids are derived from (domain, source, position, content), so re-ingesting an unchanged
document rewrites the same leaves instead of duplicating them.
It watches layer 0. When branch_factor leaves of one domain have no parent, it fires a
Summarization Event: they are condensed into a layer-1 node, the node is written, and only then
are the children stamped with their parent_id — so an interrupted run re-processes a group instead
of orphaning it. The same rule applies from layer 1 to layer 2, and so on, until the tree converges;
then root.json is rewritten.
consolidate() is idempotent. flush() (--force) also promotes partial groups so the tree closes
completely. Anything still unconsolidated is referenced directly by the trunk, so nothing in the
store is ever unreachable from the root.
Ranking uses Okapi BM25 over each candidate's title, keywords, tags and excerpt, with field weights
and document frequencies computed across the siblings of the current level — exactly the
discrimination the descent needs, at no extra I/O cost. The walk keeps a beam of beam_width
branches per level.
Confidence blends two independent signals:
confidence = 0.6 × coverage + 0.4 × margin
Coverage asks "does this leaf actually contain what was asked?"; margin asks "is it distinguishable from its rivals?". A common first name scores high on the first and near zero on the second — which is precisely when the system must not guess:
$ python main.py ask "Luis"
"Luis" is ambiguous (3 memories match this query equally well, confidence 0.62).
It could belong to any of these branches:
1. [personal] journal_personal.txt · 2024-07-22 — Vacaciones en Benidorm con Luis, mi primo…
2. [personal] journal_personal.txt · 2026-08-30 — Fin de semana en la playa de El Saler con Luis…
3. [personal] journal_personal.txt · 2024-11-30 — Cierre de trimestre… Luis Marín revisó el pull request…
Give me a clue to narrow it down (e.g. a place, a date, a project):
> vacaciones en Benidorm
query: "Luis + vacaciones en Benidorm" · confidence 0.74 · 5 node reads, 3 leaf reads, depth 2
1. [personal] journal_personal.txt · 2024-07-22
path root → N2-personal-ea34db33ae → N1-personal-59bf8bc291 → L-personal-0002-a17b3c8f35
The clue triggers a joint cross-reference: candidates matching both term groups are boosted
(×1.6), candidates matching only one are demoted (×0.7). The loop runs up to max_clue_rounds
times, stops early if the user declines, and never silently converts an ambiguous query into a
confident answer.
clue_provider is any callable, so the loop works for a human at a terminal (input) or for an
agent that lets the LLM supply its own follow-up.
Facts are (subject, attribute, value, timestamp, source_leaf), recorded programmatically via
record_fact() or extracted at ingest time by pattern rules (vivo en X / I live in X,
mi stack es Y, trabajo en Z, Diagnóstico:, Tratamiento:, Medicación: …; extend with
add_pattern).
Each (subject, attribute) keeps a dated timeline. The newest value is ACTIVE; every earlier one
becomes HISTORICAL with superseded_by and a valid_from/ valid_to interval. Nothing is
deleted, so both questions stay answerable:
memory.resolver.active_state()["user::location"].value # 'Valencia' (now)
memory.resolver.state_at("2024-12-31")["user::location"].value # 'Madrid' (back then)
Repeating a value at a later date is a confirmation, not a contradiction. A genuine change produces a
transparent notice — "On 2024-03-11 you said user location = 'Madrid', but on 2026-02-03 you updated to
'Valencia'." — and the superseded leaf is annotated, so retrieving outdated material always arrives
with its correction attached (SearchResult.notices).
A leaf itself stays ACTIVE: what it says was true on its own date, and that remains the right
answer to a historical question. What changes is that it can no longer be read as current.
class SummarizerBackend(Protocol):
name: str
def summarize(self, children, *, domain: str, layer: int) -> NodeSummary: ...
HeuristicSummarizer(default) — stdlib extractive summarization: TF term ranking with EN/ES stop words plus representative-sentence selection. Offline, deterministic, which is what lets the test suite assert on tree shape.AnthropicSummarizer(optional) — used only when selectedand theanthropicpackage andANTHROPIC_API_KEYare both present. Every failure path (missing SDK, missing key, network error, unparseable reply) falls back to the heuristic backend, so a consolidation is never lost because a model was unreachable.
python3 init_ihmt.py --backend anthropic # model set by summarizer_model in ihmt.config.json
Any other model or local runtime plugs in by implementing the same protocol and passing it as
IHMT(..., backend=MyBackend()).
| Command | Purpose |
|---|---|
init [--branch-factor N] [--target-tokens N] [--force] |
create the store |
ingest <paths…|-> [--domain D] [--type T] [--tag X] [--no-consolidate] |
ingest files, directories or stdin |
consolidate [--force] |
run pending Summarization Events |
search <query> [--top-k N] [--full] |
walk the tree |
ask <query> [--clue TEXT] [--top-k N] |
search with the clue loop |
tree [--depth N] |
outline of the hierarchy |
stats ,facts [--subject S] ,conflicts [--subject S] |
inspection |
rebuild |
rebuild catalog, timeline and trunk from the files |
demo |
end-to-end walkthrough |
--path selects the store directory and --json emits machine-readable output; both work before or
after the subcommand.
ihmt_memory/ihmt.config.json:
| Key | Default | Meaning |
|---|---|---|
branch_factor |
8 | children per branch; the log base of retrieval cost |
target_tokens /max_tokens |
2000 / 3000 | leaf size target and oversize threshold |
beam_width |
3 | branches kept alive per level |
confidence_threshold |
0.45 | below this, ask for a clue |
ambiguity_margin |
0.18 | score gap under which candidates count as tied |
max_clue_rounds |
3 | clue-loop iterations |
summarizer_backend /summarizer_model |
heuristic /claude-sonnet-5 |
summarization |
code_chunk_mode |
pack |
symbol stores one leaf per class member (used by project indexes) |
Small corpora deserve a small branch factor — the demo uses branch_factor=4, target_tokens=400 so a
handful of documents still builds a genuine multi-layer tree.
python3 -m unittest discover -v # from the project root
187 tests — 25 of them for the MCP server, skipped without the SDK — covering: byte-exact reconstruction and boundary-depth invariants for Java and Python, scene/date/section atomicity, oversized-block handling, leaf header round-trips, catalog recovery, summarization thresholds, upward propagation, idempotence, full reachability from the root, descent cost bounds, the clue loop, recency weighting, historical preservation, the project indexes, the MCP tools, the graphical interface and the CLI.
- Retrieval is a descent, not a scan. That is the whole point, and it means a branch pruned at the trunk is not revisited. Domain-level pruning only happens when a query has actual signal at the trunk; if it has none, every domain stays in play and the beam applies one level down. The clue loop is the recovery mechanism when the descent goes wide.
- Lexical, not semantic. Matching is BM25 over accent-folded, CamelCase-split tokens: it works in
any language and needs no model, but it will not match a synonym. Plugging an embedding re-ranker
into
BM25Rankeris the natural upgrade; the tree structure does not change. - Token counts are estimated at ~4 characters per token. Budgets only need to be consistent, not exact.
- Fact extraction is pattern-based. The bundled rules cover common English/Spanish phrasings and
clinical headers;
record_fact()is the reliable path, andadd_pattern()extends the rules. - Single-writer. Writes are atomic (
tmp+os.replace) and catalog rebuilds take a lock file, but the store assumes one writer at a time. initialize(force=True)discardsderived state only — branches, catalog, timeline — and detaches the surviving leaves so the next consolidation rebuilds the hierarchy. Leaf content is never deleted.
ihmt/
api.py IHMT facade wiring everything together
config.py IHMTConfig
models.py MemoryLeaf, BranchNode, ChildRef, RootIndex, Fact, Contradiction
storage.py MemoryStore: atomic I/O, catalog, recovery
textutils.py tokenizing, keywords, entities, extractive summary, timestamps
detectors.py DomainDetector
chunkers/ base · code · narrative · temporal · process · generic
summarizers.py SummarizerBackend · Heuristic · Anthropic
universal_ingestor.py UniversalIngestor
recursive_summarizer.py RecursiveSummarizer
semantic_navigator.py SemanticNavigator, BM25Ranker, ClueRequest, scope/path filters
project_index.py ProjectIndex: per-project code cache with checksum sync
conflict_resolver.py ConflictResolver + timeline manager
ihmt_gui/ local graphical interface (stdlib only)
mcp_server.py MCP server: the nine tools
init_ihmt.py · main.py · gui.py · examples/ · tests/
GUIDE.md installation and usage guide
INSTALL.md installation instructions for AI agents
templates/ memory-instructions.md: the usage rules agents append to CLAUDE.md / AGENTS.md
MIT © 2026 Gonzalo Román Márquez (gonzaroman)