TrueMemory

---

Why TrueMemory

It finds signal in the noise. Your AI sees thousands of messages. TrueMemory figures out which ones actually matter and throws away the rest. No manual tagging, no prompt engineering. It just knows.

It gets sharper over time. It's not a static database. It's a living memory that grows with you. It resolves contradictions when you change your mind, updates stale facts, and consolidates what it knows. The longer you use it, the better it gets.

It works without you thinking about it. TrueMemory automatically captures memories from your conversations and automatically injects the right ones into your next session. You never have to store or search for anything manually. It just happens.

It's 100% local. One SQLite file on your machine. Your memories never leave your device. No cloud, no API keys needed. Your data is yours. (Anonymous usage telemetry is the only exception — never memory content — and one env var turns it off. Details.)

Without TrueMemory: "What framework are we using?" Asked for the 12th time this week. Your agent starts every session with amnesia. It doesn't know your name, your stack, or anything you told it yesterday.

With TrueMemory: Your agent already knows you use FastAPI, prefer Pydantic v2, and that your auth middleware lives in src/auth/. It remembers your corrections, your preferences, and your decisions. Across every session, forever.

---

How TrueMemory Compares

System	LoCoMo	LongMemEval	Local-first	Auto-capture	License
TrueMemory Pro	93.0%	92.0%	✅	✅	AGPL-3.0
TrueMemory Base	92.0%	84.1%	✅	✅	AGPL-3.0
Mem0	61.4%	—	Partial	❌	Apache-2.0
Supermemory	65.4%	—	❌	❌	Cloud API
MemOS	75.8%	—	✅	❌	Apache-2.0
ReadAgent	79.5%	—	❌	❌	Research

All benchmarks independently reproducible. Scripts included in benchmarks/.

---

Quick Start

For Claude Code / Claude CLI / Cursor / Codex CLI / Gemini CLI

curl -LsSf https://raw.githubusercontent.com/buildingjoshbetter/TrueMemory/main/install.sh | sh

Installs everything in an isolated environment. Downloads ~1.5GB of AI models. No data leaves your machine. No sudo required.

1. Open a terminal: Mac: Cmd + Space, type Terminal. Linux: Ctrl + Alt + T. Windows: open PowerShell.
2. Paste the command above and press Enter.
3. Wait 3-5 minutes for models to download.
4. Quit your AI tool completely and reopen it (Mac: Cmd+Q).
5. Type "Set up TrueMemory" and pick a tier.

Windows (PowerShell):

irm https://raw.githubusercontent.com/buildingjoshbetter/TrueMemory/main/install.ps1 | iex

That's it. TrueMemory remembers your conversations automatically from here. Need help? Join our Discord.

If TrueMemory saves you time, a ⭐ helps other devs find it.

For developers (Python library)

pip install truememory

from truememory import Memory

m = Memory()
m.add("Prefers dark mode and TypeScript", user_id="alex")
print(m.search("preferences", user_id="alex"))

---

Tiers

Same architecture, three tiers. All included in a single install. Switch anytime by saying "switch to Pro" or "switch to Base."

	Edge	Base	Pro
LoCoMo	89.6%	92.0%	93.0%
LongMemEval	—	—	92.0%
BEAM-1M	—	—	76.6% (SOTA)
Embedding model	8 MB lightweight	600 MB high-accuracy	600 MB high-accuracy
Reranker	22M params	149M params	149M params
HyDE search	—	—	✅ (requires LLM API key)
Runs on	Any machine, CPU only	4 GB+ RAM	4 GB+ RAM + API key

Edge works everywhere. Base is the strongest fully-offline tier. Pro adds AI-powered query expansion for the highest scores.

---

Benchmarks

Tested across three major benchmarks with all systems sharing the same answer model (GPT-4.1-mini), judge (GPT-4o-mini, 3x majority vote), and scoring pipeline.

Benchmark	What it tests	TrueMemory Pro
LoCoMo	1,540 questions across 10 conversations	93.0%
LongMemEval	500 multi-session questions	92.0%
BEAM-1M	700 questions at 1M+ tokens	76.6% (SOTA)
BEAM-10M	200 questions at 10M tokens	65.0%

Reproduce any result yourself

Every benchmark script is self-contained and runs on Modal.

• LoCoMo Scripts — 8 systems (TrueMemory, Mem0, Zep, Engram, etc.)
• LoCoMo Results — per-category breakdowns, latency, cost
• LoCoMo Eval Config — exact models, prompts, parameters
• LongMemEval Scripts — oracle + strict variants
• LongMemEval Results — 6 TM Pro runs + 5 competitor results
• BEAM-1M Script — 35 conversations at 1M+ tokens
• BEAM-10M Script — 10 conversations at 10M tokens
• BEAM Results — 3 runs (1M) + 1 run (10M)

All benchmarks use the same eval pipeline. Nothing is hidden. Full details: LoCoMo | LongMemEval | BEAM

---

Works With

Lifecycle hooks capture conversations automatically. No manual work needed. Your memories stay local in a single SQLite file.

---

Python API

from truememory import Memory

m = Memory()

m.add("Prefers dark mode and TypeScript", user_id="alex")
m.add("Works at Anthropic as a senior engineer", user_id="alex")
m.add("Always sign commits with GPG", directive=True)  # directives auto-load every session

results = m.search("What are Alex's preferences?", user_id="alex")
results = m.search_deep("career history?", user_id="alex")  # multi-round, higher accuracy

Method	Description
m.add(content, user_id)	Store a memory (directive=True for standing instructions that auto-load at the start of every session)
m.search(query, user_id)	Search (6-layer pipeline + reranker)
m.search_deep(query, user_id)	Multi-round agentic search
m.get(id) / m.get_all(user_id)	Retrieve memories
m.update(id, content) / m.delete(id)	Modify or remove
m.stats()	System statistics

Full API reference →

---

Docs

Getting Started	Install to first memory
Python API Reference	Full Memory class reference
MCP Tool Reference	All 11 MCP tools
CLI Reference	truememory-mcp and truememory-ingest
Environment Variables	All TRUEMEMORY_* config options
Architecture Deep Dive	6-layer retrieval pipeline, encoding gate
Tier Selection Guide	Edge vs Base vs Pro
Debugging	Logs, traces, common issues

---

FAQ

Everything lives locally in ~/.truememory/memories.db. Edge and Base tiers make zero external calls. Pro sends only your search query text to an LLM for query expansion. Your memories are never transmitted.

No. The installer uses uv to manage a sandboxed Python 3.12. Your system Python is never touched.

Context windows are expensive, slow, and empty at the start of every session. TrueMemory gives instant recall for zero tokens of context, in under 200ms.

Anonymous usage telemetry (tool calls, session counts, platform info) is on by default. We never track memory content, queries, file paths, or API keys. Opt out: export TRUEMEMORY_TELEMETRY=off

---

Get Started in 60 Seconds

For Claude Code, Claude CLI, Cursor, Codex CLI, or Gemini CLI:

curl -LsSf https://raw.githubusercontent.com/buildingjoshbetter/TrueMemory/main/install.sh | sh

Using it as a Python library instead? pip install truememory

Questions? Join our Discord or open a Discussion. If TrueMemory saves you time, give us a ⭐

---

Thanks to Our Contributors

---

Research

TrueMemory is backed by a research paper (arXiv preprint) on retrieval-centered agent memory.

Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall (arXiv 2605.04897)

@article{sauronlabs2026storage,
  title   = {Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall},
  author  = {Sauron Labs},
  journal = {arXiv preprint arXiv:2605.04897},
  year    = {2026},
  url     = {https://arxiv.org/abs/2605.04897}
}

---

Community

• Join our Discord for help, feedback, and updates
• Follow @Building_Josh on X for updates
• Follow @Sauron_Labs for company news
• Open a Discussion for questions or ideas
• Read the paper on arXiv
• Visit truememory.net · sauronlabs.ai

If TrueMemory saves you time, give us a ⭐

---

Translations

Read this in: 简体中文 · हिन्दी · Español · Français · العربية · বাংলা · Português · Русский · 日本語 · 한국어 · Deutsch · Bahasa Indonesia · Tiếng Việt · Türkçe · Italiano · ไทย · Polski · Українська · Nederlands

---

License

AGPL-3.0. Free for personal and research use. Commercial use requires a separate license. Contact [email protected].

---