hypermnesic

One brain. Every AI. Yours.

Your second brain lives as plain Markdown in a Git repo you host. ChatGPT, Claude, and your coding agents — on your laptop and your phone — all read and write that same brain through one endpoint. Obsidian is how you browse it.

Here's the part that matters: because every AI keeps curating the same notes, your memory doesn't just persist — it compounds. Every fact captured, every correction, every link added makes the next answer from every assistant sharper. Per-app memory traps you in silos; one shared brain gets smarter each day — and it's plain files you own, not rows in a vendor's database.

Every memory is a real Git commit — reviewable, revertible, yours. The search index is a disposable projection of your files; delete it and rebuild it any time. A reindex can never lose a memory.

Who it's for: developers and power note-takers who want durable agent memory they own — plain files in their own Git history, reachable by every assistant, with no vendor lock-in.

Status: public v0.1.0 release. The engine is licensed under AGPL-3.0-only; the companion plugin ships from the separate GPL-3.0 hypermnesic-companion repository.

Why it compounds

Per-app memory fragments: what you told ChatGPT is invisible to Claude, your phone's assistant forgets what your laptop's agent learned, and none of it is yours to move. A shared brain does the opposite — it turns every interaction into a flywheel:

• Capture — any AI writes a note: a decision, a fact, a person, a meeting.
• Curate — any AI links it, corrects it, or adds context the next time it's relevant.
• Recall — every other AI retrieves it on the next question.
• Compound — the brain grows denser and more useful with every turn, for all of them.

Because every write is a reviewable Git commit, one assistant's curation is safe, visible, and revertible for all the others. The brain is shared and trustworthy.

One endpoint, every client

Point any MCP-capable app at your endpoint URL and it just works — OAuth is automatic (log in through the browser once, then silent refresh):

• ChatGPT, Claude (desktop, mobile, web) — add a custom connector.
• Claude Code / Codex — the bundled plugin.
• Your own agents — any MCP client, same URL.
• Obsidian — a read-only companion over your tailnet, for browsing and serendipity.

On the machine that holds the vault, skip the network entirely and use the hypermnesic CLI. Setup details are in the Quick start below.

How it's different

Most "agent memory" keeps your memories in their store. hypermnesic keeps them as plain Markdown in your Git repo; the search index is a throwaway projection you can delete and rebuild at will. Everything else follows from that one choice.

How it sits next to tools you may know:

• mem0 / Zep — memory APIs over a managed vector (and graph) store. Reach for them if you want a hosted memory service; reach for hypermnesic if you want your files to be the memory.
• Hindsight — also open-source agent memory, but it lives in its own vector store you run via Docker/cloud, and it posts a higher LongMemEval score on a more lenient judge axis. hypermnesic optimizes for owned, auditable, compounding files, not a leaderboard rank — read the honest comparability envelope in harness/BENCHMARKS.md.
• Honcho — complementary, not competing. Honcho models who you are (preferences, style, theory-of-mind); hypermnesic holds what you know, in files. Use both.
• A database-backed personal brain — I built one before this. The database drifted from the files, and I couldn't fully trust or move it. hypermnesic is the rebuild: files are truth, the index is disposable.

Full tool-by-tool detail — including when hypermnesic is the wrong fit — is in why hypermnesic.

Try it in under 5 minutes

uv tool install hypermnesic
hypermnesic local-proof --demo-dir /tmp/hypermnesic-demo

That creates a tiny Markdown git repo, projects it into the disposable index, recalls the repo-relative source note, and previews the exact commit_note write diff without writing it. No account, no service.

The demo source is an asciinema cast. It uses only a generated /tmp/hypermnesic-demo vault and placeholder-safe paths.

Quick start

A. Prove local memory works

Start on the machine that holds the vault. You need only a git repo of markdown notes. Dense embeddings improve ranking when OPENAI_API_KEY is configured, but the proof also works offline in lexical mode.

# 1. install the engine from PyPI
uv tool install hypermnesic

# 2. prove recall from your own markdown files, with a dry-run write preview
hypermnesic local-proof /path/to/your/vault

# or try a tiny generated demo vault first
hypermnesic local-proof --demo-dir /tmp/hypermnesic-demo

The proof path validates a git-backed vault, projects committed markdown files into the disposable .hypermnesic/ index, asks a natural-language question, returns the repo-relative source markdown path, and shows a commit_note dry-run diff without creating a write commit. The success milestone is Local memory works.

B. Self-host the endpoint

After the local proof succeeds, bring the shared endpoint online for remote apps. You need Tailscale installed and logged in (tailscale up); hypermnesic uses Tailscale Funnel for public HTTPS + automatic TLS, so there is no reverse proxy or cert to manage.

hypermnesic setup /path/to/your/vault \
  --public-url https://<your-host>.ts.net/mcp

# later, diagnose without changing services, secrets, funnel routes, or git
hypermnesic doctor /path/to/your/vault \
  --public-url https://<your-host>.ts.net/mcp

setup renders + starts a user service, generates an owner-only consent secret (~/.config/hypermnesic-cloud/cloud.env, chmod 600), configures the Tailscale funnel (the /mcp mount + the OAuth discovery well-knowns), then verifies the live HTTPS discovery chain before reporting success. Re-running it converges to the same state. It prints milestone checks, your endpoint URL, and login instructions. --resource defaults to --public-url; pass it only when the OAuth resource identifier differs. doctor and status report local index health, remote reach, OAuth discovery, auth challenge, write availability, dense key source/state, vector coverage, and client-specific next actions without mutating state. Read tools also expose degraded_reason; after an embedding 429 they cool down provider calls while continuing to serve lexical/graph results. Dense key lookup is repo-scoped: process OPENAI_API_KEY wins, otherwise the target vault's gitignored .env is used even when the command or MCP server starts from another working directory. Live OpenAI smoke checks stay opt-in with --check-dense-live. Dense diagnostics keep credential state separate from projection state: missing indexes point to initialization, and stale/absent vectors point to hypermnesic converge /path/to/vault --now --json before a full reindex. By default, new OAuth clients request read; admins can make new approvals request both read and write with --default-client-scopes read write or HYPERMNESIC_DEFAULT_CLIENT_SCOPES=read,write.

C. Connect a client (any remote app)

Point the app's MCP server at your endpoint URL — that's it. OAuth is automatic: