retainr 0.4.0

Persistent, queryable AI memory for any Python app — local-first, zero API cost

retainr

Persistent, queryable memory for any Python AI app — local-first, zero API cost.

---

What is it?

retainr gives any AI app a long-term memory layer — without needing an API key, a cloud service, or a database server.

Store things. Recall them semantically. Forget them. All locally.

from retainr import Memory

mem = Memory(user_id="arjun")

mem.remember("I prefer dark mode and use VS Code")
mem.remember("Currently building a job application assistant")
mem.remember("I like concise answers over long explanations")

results = mem.recall("what tools does the user prefer?")
for r in results:
    print(f"[{r['score']:.3f}] {r['text']}")

[0.412] I prefer dark mode and use VS Code
[0.289] I like concise answers over long explanations

CLI Demo

Stats

---

Install

pip install retainr

---

Features

• 🧠 Semantic recall — finds memories by meaning, not exact keywords
• 📉 Memory decay — old memories fade in relevance automatically
• ⭐ Importance scoring — weight memories by how significant they are
• 🏷️ Tags — organize memories into namespaces
• 🗑️ Forget & clear — GDPR-friendly deletion
• 💾 Local-first — uses FAISS + SQLite, no external services
• 🆓 Zero cost — no API keys, no cloud, runs on your machine

---

API Reference

Memory(user_id, db_path)

from retainr import Memory

mem = Memory(user_id="alice", db_path="memory.db")

Param	Default	Description
user_id	"default"	Isolates memories per user
db_path	"memory.db"	SQLite file path. Use ":memory:" for in-RAM (tests)

---

remember(text, tags, importance, decay_days)

Store a memory.

mid = mem.remember(
    "User just got promoted to SDE-2",
    tags=["career", "milestone"],
    importance=0.9,   # 0.0–1.0, default 0.5
    decay_days=90,    # score halves every 90 days, 0 = no decay
)

Returns: str — the memory ID (use it to forget() later)

---

recall(query, top_k)

Find the most semantically relevant memories.

results = mem.recall("what is the user's job status?", top_k=3)
# [{"id": ..., "text": ..., "score": 0.87, "tags": [...], ...}]

---

forget(memory_id)

Delete a specific memory.

mem.forget(mid)

---

clear()

Wipe all memories for this user.

mem.clear()

---

export(filepath)

Backup all memories to JSON.

mem.export("my_memories.json")

---

How it works

1. remember() → text is embedded using sentence-transformers (all-MiniLM-L6-v2, runs locally) → vector stored in FAISS, metadata in SQLite
2. recall() → query is embedded → FAISS finds top-k nearest vectors → SQLite returns the original text
3. Scoring → final score = cosine similarity × (0.7 + 0.3 × decayed importance)

---

Why not mem0 / Zep / etc.?

	retainr	mem0	Zep
Local-first	✅	❌	❌
API key needed	❌	✅	✅
Cost	Free	Paid tiers	Paid tiers
Memory decay	✅	❌	❌
pip install	✅	✅	✅

---

Stack

• sentence-transformers — local text embeddings
• faiss-cpu — vector similarity search
• sqlite3 — metadata storage (built into Python)

---

License

MIT © Devarajan