supamem 0.3.0a1

Project-agnostic dual-memory tooling for Claude Code, Cursor, and opencode

Languages: English · 简体中文 · Español · 日本語 · Русский

🧠 supamem

Qdrant-backed dual-memory for AI coding agents

Give Claude Code, Cursor, and OpenCode persistent semantic + structural memory across every project.

---

👋 Built by Dzmitry Sukhau — AI-native Solution / Software Architect / CTO

Available for consulting on AI products, integrating AI into existing products, and business-process automation.

If you're shipping LLM features, evaluating retrieval pipelines, hardening agentic systems, or building an AI-first product from scratch — let's talk.

  

---

✨ What is supamem?

supamem is a single-binary CLI that wires up a production-grade memory layer for any AI coding assistant. Drop it into a fresh repo, run supamem init, and your agents instantly gain:

• 🔍 Semantic search over project notes, ADRs, decisions, and past conversations (hybrid sparse+dense retrieval)
• 🤖 MCP server that any compatible client (Claude Code, Cursor, OpenCode) can talk to
• 🪝 Per-client hooks that auto-load relevant memory at session start and on file edits
• 📊 Welford usage stats so you can see what memory is actually being recalled
• 🧪 Eval harness with a 33-query golden corpus to detect retrieval regressions

Battle-tested inside SoftChat (Phases 80.1–80.5) before being extracted into a standalone package every team can adopt.

---

🎯 Why supamem exists

The problem: Coding agents have no memory between sessions. Every time you open a new conversation in Claude Code / Cursor / OpenCode, the model has zero context about your codebase, past decisions, ADRs, known issues, or conventions. So either:

1. You re-paste 5–15 KB of context at the start of every session (slow, error-prone, costly), or
2. You let the agent flounder — it grep-walks the repo, asks redundant questions, forgets last week's decisions, and rediscovers the same gotchas you already documented six months ago.

The fix: A persistent semantic + structural memory layer that automatically retrieves the right 1–2 KB of context for the current prompt — no manual pasting, no re-explaining, no context blow-out.

Phase 80.1 bench (33 labeled goldens, real Claude Code sessions): −78.5% tokens vs naive whole-doc retrieval at the same recall, p95 73 ms end-to-end.

The full evaluation is the same one we ran inside SoftChat to lock the production pipeline. Methodology: 33 representative dev queries → 4 retrieval arms compared (baseline_union, tuned_current, tuned_hybrid, mem0_vector) → token count + recall CI + latency measured per arm.

📊 Token consumption: agent with memory vs without

Numbers below are per typical 30-turn Claude Code session assuming a real codebase with ~50 ADRs / insights / rules (≈ what SoftChat ships). YMMV — but the ratio between arms holds.

Approach	Tokens/turn	Tokens/30-turn session	Notes
❌ No memory layer	≈ 0 auto-injected, but you paste context manually	30,000–80,000 (manual paste, repeated)	You spend cognitive load on copying instead of building
⚠️ Naive RAG (whole-doc embed)	~5,800 / turn	~174,000	Bloated, recalls big files when you only needed a paragraph
✅ supamem tuned_hybrid	~1,250 / turn	~37,500	Same recall, −78.5% tokens vs naive RAG

💰 Approximate inference cost savings

Anthropic API list pricing (Mar 2026): Sonnet 4.6 = $3 / Mtok input · Opus 4.7 = $15 / Mtok input.

Model	Tokens saved/session vs naive RAG	Cost saved/session	Monthly (110 sessions)
Sonnet 4.6	136,500	$0.41	~$45/dev
Opus 4.7	136,500	$2.05	~$225/dev

A 10-engineer team running Opus saves ~$2,250/month on input tokens alone — without counting the cost of slower iteration, lost decisions, and time spent re-pasting context. Output token savings (less hallucination, fewer back-and-forth turns) compound on top.

🥊 vs the alternatives

	No memory	Naive RAG	mem0 / atomic facts	supamem (tuned_hybrid)
Auto-inject on session start	❌	⚠️	✅	✅
Hybrid sparse+dense retrieval	❌	❌	❌	✅
Code-identifier preservation	❌	✅	❌ (drops names)	✅
Locked schema + golden eval	❌	❌	❌	✅
Multi-client (Claude/Cursor/OpenCode)	❌	❌	⚠️	✅
p95 latency	n/a	~120 ms	~80 ms	73 ms
Token bloat	High (manual)	Highest	Low but lossy	Lowest with full recall

Why hybrid? BM25 catches exact identifiers (ChatService.generate, env-var names, file paths) that dense embeddings smear. Dense catches semantic intent ("how do we handle billing webhooks?") that BM25 misses. RRF fusion combines both rankings so you get the best of each.

Why not mem0? mem0's atomic-fact extraction loses code identifiers — recall on the 33-query bench was 0.015 (effectively zero). Great for personal CRM-style memory, not for code-aware retrieval.

---

⚡️ 60-second quickstart

# 1. Install (uv is the fastest path)
uv tool install supamem

# 2. Start Qdrant (one-time, ~30s)
docker run -d -p 6333:6333 -p 6334:6334 -v $HOME/.qdrant:/qdrant/storage qdrant/qdrant:latest

# 3. Bootstrap your project
cd your-project
supamem init

# 4. Wire it into your AI client
supamem install --client claude-code   # or cursor, opencode

# 5. Confirm everything is healthy
supamem doctor

That's it. Open Claude Code (or your preferred client) inside the project — the memory tool is already on the menu. ✨

---

👀 See it work — supamem live

Run supamem live in a side terminal to watch every retrieval call as it happens — perfect alongside Claude Code / Cursor / OpenCode for instant visibility into the silent PreToolUse-hook injections (which save tokens by NOT showing UI).

The SessionStart banner (v0.1.4+) also lands a one-line status in your AI client at session open: 🧠 supamem v0.1.4 · <collection> · <N> chunks · audit <path> — auto-detects Claude Code / Cursor / OpenCode via env vars.

🎬 Interactive demo: supamem-live.cast — drop into asciinema.org/player or run locally with asciinema play docs/media/supamem-live.cast.

---

🚀 Features

Feature	Description
🔍 Hybrid retrieval	Tuned sparse (BM25) + dense (MiniLM) fusion, locked schema D-25
🎯 Code-aware reranker	Cross-encoder mxbai-rerank-base-v2 (Apache-2.0) rescores tuned_hybrid candidates by default. Disable with retrieval.reranker = "off" for pre-v0.2.4a1 behavior. (Phase 8, RERANK-01..04)
⏳ Per-source temporal validity	Every chunk carries valid_from/valid_to; re-indexing a changed file supersedes prior chunks atomically and the retrieval-time filter excludes superseded points across every backend. Optional transcript-only recency decay (off by default). Auto-GC past retention_days = 90 (set to 0 for kept-forever / audit collections). (Phase 9, TEMP-01..03)
📚 Markdown chunker	Header-aware, 200-token chunks with 250-token soft max (T-1)
🤖 MCP server	stdio (default) and http transports, official mcp SDK
🪝 Multi-client hooks	Claude Code session-start, OpenCode session-start, Cursor MDC
🧰 One-command install	Atomic config patching with auto-backup and rollback
🩺 supamem doctor	Probe Qdrant, resolve config chain, surface version drift
👀 supamem live	Rich-Live terminal dashboard tailing the audit JSONL — real-time visibility into retrieval calls (v0.1.4+)
🎬 SessionStart banner	One-line cross-client banner injected at session open (Claude Code / Cursor / OpenCode), v0.1.4+
📊 Welford counters	Track recall rate, latency, query volume per project
🧪 Eval harness	33-query golden corpus + regression detection
🔁 Brownfield migration	Detect existing dev_memory and migrate non-destructively
🎨 Stylish CLI	Rich-powered spinners, panels, and color so you always see progress

---

📋 Prerequisites

You only really need two things: Python 3.12+ and Qdrant. Everything else is optional.

🐍 Python 3.12+  ·  click to expand install commands

# macOS (Homebrew)
brew install python@3.12

# Linux (Ubuntu/Debian)
sudo apt install python3.12 python3.12-venv

# Windows (PowerShell)
winget install Python.Python.3.12

We strongly recommend installing uv — the fastest Python package manager:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

🗄️ Qdrant 1.10+  ·  vector database (required)

The simplest path is Docker:

docker run -d --name qdrant \
  -p 6333:6333 -p 6334:6334 \
  -v $HOME/.qdrant:/qdrant/storage \
  qdrant/qdrant:latest

Or with docker compose:

services:
  qdrant:
    image: qdrant/qdrant:latest
    ports: ["6333:6333", "6334:6334"]
    volumes: ["./qdrant_data:/qdrant/storage"]
    restart: unless-stopped

Don't have Docker? Run a managed cluster on Qdrant Cloud (free tier available) and point supamem at the URL via supamem init.

🤖 An MCP-compatible client  ·  pick at least one

Client	Install	Notes
Claude Code	npm install -g @anthropic-ai/claude-code	First-class MCP support
Cursor	Download from cursor.com	Uses MDC rules + MCP
OpenCode	curl -fsSL https://opencode.ai/install | bash	Open-source TUI, MCP native

---

📦 Install

# Recommended: uv (fastest, isolated)
uv tool install supamem

# Alternative: pipx (also isolated)
pipx install supamem

# Plain pip (in a venv)
pip install supamem

Verify:

supamem --version

You should see a colorful banner and the credit line. 🎨

Latest: v0.1.4 is published on PyPI. Released via Trusted Publisher OIDC — every wheel is provenance-attested.

Models cached at install

supamem install <client> and supamem init proactively download all ML prerequisites (MiniLM ~90 MB, BM25 ~10 MB, mxbai-rerank-base-v2 ~1 GB) with a progress bar. Cold post-install CLI invocations (supamem --help, supamem doctor, supamem --version) trigger zero network egress. Air-gapped first-run? Pass --skip-models, then run supamem repair once network is available.

Models live under platformdirs.user_cache_dir("supamem")/models/ (override with SUPAMEM_CACHE_DIR).

Subagent reachability (v0.2.5+)

If you use Claude Code subagents shipped by GSD, superpowers, hookify, or any plugin that pins a tools: whitelist on its agent definitions, those agents cannot reach the supamem MCP server unless mcp__supamem__* is in the whitelist — even when the parent session has supamem connected. Subagents inherit only the tools their frontmatter lists.

supamem install and supamem repair patch this for you automatically:

supamem install --client claude-code   # patches ~/.claude/agents/ + <project>/.claude/agents/
supamem repair                         # re-applies if a plugin overwrites your agents

The patcher is idempotent (running twice produces zero changes), preserves your YAML style (CSV vs list), and skips symlinked agent files with a warning. Files with a missing or empty tools: line have full inheritance per Claude Code semantics and are left untouched.

Backup manifest lives at ~/.cache/supamem/agent_patches.json. Reverse cleanly with:

supamem unpatch-agents

Pass --skip-patch-agents to opt out on any of install / init / repair.

Uninstalling supamem

supamem unpatch-agents      # restore agent whitelists first
pip uninstall supamem

There is no portable pip / uv / pipx uninstall hook in 2026, so the two-step is the supported contract. supamem doctor shows the manifest path and reminder so you can discover this flow naturally.

---

🎯 CLI surface

Command	Purpose
supamem init	Greenfield bootstrap — probes Qdrant, creates collection, writes .supamem/config.toml
supamem install --client <name>	Patch a client config (claude-code, cursor, opencode) — atomic with backup. Defaults to --scope project (per-workspace files); pass --scope user for legacy global behavior. Pass --enforce-search (claude-code only) to wire the opt-in edit-gate hook. v0.2.5+: auto-patches ~/.claude/agents/ and <project>/.claude/agents/ to add mcp__supamem__* to restrictive tools: whitelists; opt out with --skip-patch-agents.
supamem repair	🩹 Migrate from legacy global install to per-workspace files. Strips stale mcpServers.supamem from globals and re-installs at project scope from the current cwd. v0.2.5+: re-applies subagent reachability patches. Idempotent. Supports --skip-patch-agents.
supamem index	Embed dev memories into Qdrant using the locked tuned-hybrid pipeline (D-25)
supamem mcp-server	Run the MCP server (--transport stdio default; --transport http for HTTP)
supamem hook <client>	Per-client session/edit hooks (called by the client itself)
supamem doctor	🩺 Probe Qdrant, print resolved config chain, report version drift
supamem stats	Welford schema-v2 usage counters from .supamem/state/
supamem live	👀 Live dashboard tailing the audit JSONL — pipe-safe (plain JSONL when not a TTY); handles rotation, resize, Ctrl-C
supamem migrate	Brownfield migration from a pre-existing dev_memory collection
supamem eval	Run the regression harness against the bundled 33-query golden corpus
supamem uninstall --client <name>	Reverse supamem install cleanly. Strips supamem from BOTH project and user scopes.
supamem unpatch-agents	🔄 Reverse subagent reachability patches (v0.2.5+). Restores agent files to their pre-patch form per the manifest at ~/.cache/supamem/agent_patches.json. Skips files you've edited since with a per-file warning. Run BEFORE pip uninstall supamem for a clean uninstall.

Environment variables

Var	Purpose
SUPAMEM_PROJECT_ROOT	Absolute path to the workspace. Honored first by mcp-server for project resolution; injected automatically by supamem install --scope project so MCP hosts that launch the subprocess from the wrong cwd still resolve the right collection.
SUPAMEM_CONFIG	Explicit TOML path overriding all discovery. Highest precedence.
SUPAMEM_GATE_DISABLE=1	Bypass the opt-in claude-code edit-gate for the current session (--enforce-search users only).
SUPAMEM_ADVISORY_DISABLE=1	Suppress the Cursor beforeSubmitPrompt advisory hook.
SUPAMEM_NO_UPDATE_CHECK=1, NO_UPDATE_NOTIFIER=1, CI=1	Suppress the GitHub Releases probe.
SUPAMEM_BANNER_DISABLE=1	Suppress the SessionStart one-line banner entirely (no context injection, no user-visible status).
SUPAMEM_BANNER_QUIET=1	Suppress only the user-visible terminal status line; keep injecting the banner into Claude Code's additionalContext for the model. Use this when you want supamem context loaded but no per-session SessionStart:supamem says: … row in your terminal.

SessionStart banner format

Every supported client emits a one-line status at session open:

🧠 supamem ✓ v0.2.0 · supamem-myproject · 412 chunks · audit /home/me/.cache/supamem/audit.jsonl
          ^── health flag (✓ healthy / ⚠ misconfigured or qdrant unreachable)

When a newer release is locally cached by the background update probe, an update v0.X.Y available segment is appended. Healing is never automatic — the banner only signals; run supamem repair to act.

Every long-running command shows a live spinner with elapsed time so you always know it's working. Use --help on any subcommand for details.

---

📜 Transcript ingestion (v0.2.2a1+)

supamem can index your Claude Code session history as Q+A drawer chunks alongside your project's Markdown corpus, surfacing past decisions and tool-use traces in dual_memory_search. Default-OFF — opt in with --transcripts.

# Index Claude Code transcripts from the default location (~/.claude/projects/)
supamem index --transcripts

# Or point at a specific directory
supamem index --transcripts /path/to/sessions/

# Skip the regular project corpus and only index transcripts
supamem index --transcripts --transcripts-only

# Limit to recent sessions (default: 180 days; --since 0 disables the filter)
supamem index --transcripts --since 30d

Configure under [supamem.transcript] in .supamem/config.toml:

[supamem.transcript]
default_root           = "~/.claude/projects/"
since_days             = 180
tool_payload_max_chars = 2000
chunk_soft_max_tokens  = 600
include_paths_glob     = []
exclude_paths_glob     = []   # exclude sensitive sessions, e.g. ["**/banking-*.jsonl"]

⚠ Transcripts may contain secrets. API keys, tokens, and other credentials sometimes end up pasted into Claude Code sessions. v0.2.2a1 ships no redaction — review your ~/.cache/supamem Qdrant collection before sharing it. Hand-exclude sensitive sessions via exclude_paths_glob. Redaction is tracked for v0.3 via a future supamem.redactor plugin group.

Currently supported transcript formats: Claude Code JSONL (Cursor SQLite and ChatGPT export are deferred to follow-on plugins).

---

🔎 Scoped retrieval (v0.2.3a1+)

Filter retrieval by coding-shaped category via the where parameter on dual_memory_search (and the qdrant_find alias):

# Only chunks classified as backend code
dual_memory_search(query="auth flow", where={"room": "backend"})

# OR across rooms (Qdrant MatchAny)
dual_memory_search(query="rate limit", where={"room": ["backend", "tests"]})

Every indexed chunk carries payload.room — one of backend, frontend, tests, docs, scripts, config, migrations, types, or null. Classification is exact path-component equality (split on /) — a file at data/chest_xray/img.png is NEVER classified as tests. Multiple keys in where are AND; list values within a key are OR.

Override the default keyword map in .supamem/config.toml:

[supamem.classifier.rooms]
tests      = ["tests", "test", "__tests__"]
backend    = ["src", "backend", "api"]
frontend   = ["frontend", "web", "client", "components"]
# Priority is encoded by key order — first match wins.
# Putting `tests` before `backend` makes tests/backend/api_test.py classify as `tests`.

supamem doctor surfaces the active rooms map with [source: ...] provenance, the stored classifier_hash, and a per-room histogram (including a null bucket).

Changing [supamem.classifier.rooms] triggers a one-time re-classify sweep on the next supamem index — Qdrant set_payload per-room, zero re-embedding cost. Pre-v0.2.3 collections auto-migrate on first post-upgrade index invocation.

Transcript chunks (chunker == transcript) classify to room = null by construction — filter them via the existing payload.chunker key.

---

🎯 Code-aware reranker (v0.2.4a1+)

Every tuned_hybrid query now rescores RRF-fused candidates through a cross-encoder (mixedbread-ai/mxbai-rerank-base-v2, Apache-2.0, ~1 GB) by default. Sharper precision on code-shaped queries; the v0.2.0 escape hatch is retrieval.reranker = "off", which restores pre-Phase-8 byte-identical behavior.

[supamem.retrieval]
reranker = "mxbai_v2"  # default in v0.2.4a1+; "off" restores pre-Phase-8 behavior

[supamem.retrieval.reranker]
model_id         = "mixedbread-ai/mxbai-rerank-base-v2"
top_n            = 50   # rerank pool size; clamps to fused-candidate count
prefetch_per_arm = 50   # widened from default 20 when reranker is on
batch_size       = 16

When the reranker is on, tuned_hybrid widens PREFETCH_LIMIT to 50 per arm, skips the T-4 recency multiplier (cross-encoder + recency-prior is anti-aligned for code retrieval per PROJECT.md), and runs T-5 cosine-dedup + T-8 token-budget AFTER rerank. RetrievedChunk.rerank_score carries the cross-encoder logit when reranker is on; the primary score is replaced by it.

supamem doctor adds a Reranker panel after the existing Retrieval panel: active reranker name, model_id, cache path, on-disk size + partial-download detection, last-load latency, last-100-query rerank p50/p95, and detected device (cuda/mps/cpu). When the cache is partial or corrupted, run supamem repair — the canonical doctor-driven self-heal entry point that re-fetches missing model files, re-syncs share/, repairs managed CLAUDE.md/AGENTS.md blocks, and restores client config. Idempotent.

Third parties register custom rerankers via the new supamem.reranker plugin entry-point group (4th group alongside retrieval / embedder / chunker):

[project.entry-points."supamem.reranker"]
my_reranker = "my_pkg.module:MyReranker"

Plugin protocol: rerank(query: str, candidates: list[RetrievedChunk]) -> list[RetrievedChunk]. Lazy model-load on first call; eager warm-up runs through the install/init/repair fetch pipeline.

---

⏳ Per-source temporal validity (v0.3.0a1+)

Every indexed chunk carries a binary valid_to field:

• valid_to = null → live
• valid_to ≤ now() → superseded (filtered out of every retrieval)

When a file changes and you re-index, the indexer atomically:

1. Scrolls every existing chunk for that file path.
2. Sets valid_to = now() on each (closes the prior validity window).
3. Upserts the new chunks under content-hash-keyed UUIDs with valid_to = null.

Old and new chunks coexist in Qdrant; only the new ones are returned by retrieval until the auto-GC sweep deletes the old ones past retention_days. The retrieval-time filter is constructed at a single site and inherited by every backend (tuned_hybrid both Prefetch arms, dense, bm25, qdrant_find, dual_memory_search) — uses Qdrant's IsEmptyCondition on valid_to (NOT IsNullCondition — see Qdrant#5342: IsNull does not match missing fields).

Configure in .supamem/config.toml:

[supamem.retrieval.temporal]
retention_days = 90          # 0 = kept forever (compliance / audit collections)

Transcript-only recency decay (opt-in, default OFF)

Code, ADRs, and docs do not "go stale". Transcripts often do — older support-chat turns with deprecated APIs distract the agent from the current dialogue. Phase 9 ships an opt-in multiplicative-floor decay knob that runs only on transcript chunks, after rerank, never auto-enabled for code / ADR / doc:

[supamem.retrieval.recency.per_source.transcript]
enabled        = true            # default false
half_life_days = 14.0
alpha          = 0.7             # floor: oldest transcript still gets 0.7x its score

Worked example with the locked defaults (alpha = 0.7, half_life_days = 14):

Age (days)	Multiplier
0	1.000
7	0.924
14	0.850
28	0.775
∞	0.700 (floor at α)

Code / ADR / doc rankings stay byte-identical when the knob is flipped — verified by an end-to-end byte-identity test (TEMP-03 success criterion).

References: Customers.ai recency-weighted scoring, Snowflake Cortex Search scoring docs.

Doctor surface

supamem doctor shows a Temporal validity panel (between Reranker and Subagent reachability) listing live / superseded / awaiting_gc / future-dated counts, per-source breakdown, oldest + newest valid_from across your collection, and retention_days provenance. Read-only by construction; never flips the doctor exit code.

Migration

First post-upgrade supamem index back-fills valid_to=null on legacy points (gated by a manifest reserved key, idempotent on subsequent runs). Defense-in-depth alongside the IsEmpty runtime filter.

⚠ Default retention is destructive for users upgrading from v0.2.x with audit-mode collections older than 90 days. Set [supamem.retrieval.temporal] retention_days = 0 to disable auto-GC entirely.

---

🪛 Wiring into your client

Claude Code

supamem install --client claude-code              # default: --scope project (per-workspace .mcp.json)
supamem install --client claude-code --scope user  # legacy global install in ~/.claude.json
supamem install --client claude-code --enforce-search  # also register the opt-in edit-gate

Default writes <repo>/.mcp.json (project-scope, committable; takes precedence over user-scope per Anthropic MCP docs). Always registers the SessionStart banner + injection hook in ~/.claude/settings.json. With --enforce-search, also registers a PreToolUse gate that DENIES Edit|Write|MultiEdit when no mcp__supamem__dual_memory_search is found in the session transcript since the last user turn — override per-session with SUPAMEM_GATE_DISABLE=1. Preview any command with --dry-run.

Cursor

supamem install --client cursor              # default: --scope project (<repo>/.cursor/mcp.json)
supamem install --client cursor --scope user  # legacy global install in ~/.cursor/mcp.json

Default writes <repo>/.cursor/mcp.json (per-workspace; project-level wins on conflict per Cursor docs). Always writes <repo>/.cursor/rules/dual-memory.mdc and registers a sessionStart snapshot hook + a beforeSubmitPrompt advisory in <repo>/.cursor/hooks.json. The advisory injects an agentMessage reminder when the user's prompt looks edit-bound; suppress with SUPAMEM_ADVISORY_DISABLE=1. (Cursor's hooks API doesn't yet support a fail-closed pre-edit event — the advisory is the strongest available nudge.)

OpenCode

supamem install --client opencode

Updates ~/.config/opencode/opencode.json and writes a session-start hook to ~/.config/opencode/hooks/.

🛟 MCP launched from the wrong cwd? Hosts (Cursor, some IDE wrappers) sometimes spawn the MCP subprocess from $HOME instead of the workspace, causing supamem to fall back to the default collection (dev_memory_tuned_hybrid) and return Qdrant 404s. Set SUPAMEM_PROJECT_ROOT=/abs/path/to/workspace in the host's MCP config (e.g. ~/.cursor/mcp.json env block, or ~/.claude.json under mcpServers.supamem.env). If unset, supamem will walk parents looking for .supamem/config.toml or pyproject.toml [tool.supamem] — and emit a one-line stderr warning when it can't find either. Verify with supamem doctor from the repo root: the resolved collection should match what your MCP client returns from dual_memory_search.

---

🧠 How it works

┌─────────────────┐    MCP/stdio     ┌─────────────────┐    REST    ┌─────────────┐
│ Claude / Cursor │ ───────────────► │  supamem MCP    │ ─────────► │   Qdrant    │
│   / OpenCode    │ ◄─────────────── │     server      │ ◄───────── │  (vectors)  │
└─────────────────┘                  └─────────────────┘            └─────────────┘
        │                                    ▲
        │ session-start hook                 │ tuned-hybrid retrieval
        ▼                                    │ (BM25 + MiniLM fusion)
┌─────────────────┐                          │
│ supamem hook    │ ─────────────────────────┘
│  (auto-recall)  │
└─────────────────┘

• Indexer chunks Markdown by header (T-1 chunker, 200-token target / 250 soft max)
• Embedders produce sparse (BM25) and dense (MiniLM-L6) vectors
• Retrieval runs both arms in parallel, fuses with reciprocal rank fusion, returns top-k
• MCP server exposes dual_memory_search (read) and dual_memory_write (write/idempotent agent-memory persistence) — plus qdrant_find and qdrant_store as drop-in aliases for users coming from upstream mcp-server-qdrant (disable with SUPAMEM_QDRANT_ALIASES=0)
• Hooks call supamem hook <client> at the right moment, so memory loads transparently

---

🤝 Contributing

We welcome PRs! Quick start:

git clone https://github.com/dzmitrys-dev/supamem.git
cd supamem
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
pytest
ruff check .

Coming from an in-tree dev_memory setup? See MIGRATION.md.

---

📜 License

MIT — see LICENSE.

---

💜 Delivered with care by

SoftChat  ·  SoftSkillz

Russian-language AI chat platform  ·  AI-first product engineering

supamem was extracted from SoftChat's production memory stack so every team can run on the same battle-tested pipeline. If it makes your agents smarter, give us a ⭐ — and check out what we build with it.

_{Made with care in Belarus  🇧🇾  ·  app.softchat.ru  ·  softskillz.ai}