Open Source · AGPL-3.0

███╗   ███╗███████╗███╗   ███╗ ██████╗ ██████╗ ██╗   ██╗
████╗ ████║██╔════╝████╗ ████║██╔═══██╗██╔══██╗╚██╗ ██╔╝
██╔████╔██║█████╗  ██╔████╔██║██║   ██║██████╔╝ ╚████╔╝
██║╚██╔╝██║██╔══╝  ██║╚██╔╝██║██║   ██║██╔══██╗  ╚██╔╝
██║ ╚═╝ ██║███████╗██║ ╚═╝ ██║╚██████╔╝██║  ██║   ██║
╚═╝     ╚═╝╚══════╝╚═╝     ╚═╝ ╚═════╝ ╚═╝  ╚═╝   ╚═╝
████████╗██╗  ██╗ █████╗ ████████╗  ████████╗██╗  ██╗██╗███╗   ██╗██╗  ██╗███████╗
╚══██╔══╝██║  ██║██╔══██╗╚══██╔══╝  ╚══██╔══╝██║  ██║██║████╗  ██║██║ ██╔╝██╔════╝
   ██║   ████████║███████║   ██║        ██║   ████████║██║██╔██╗ ██║█████╔╝ ███████╗
   ██║   ██╔══██║██╔══██║   ██║        ██║   ██╔══██║██║██║╚██╗██║██╔═██╗ ╚════██║
   ██║   ██║  ██║██║  ██║   ██║        ██║   ██║  ██║██║██║ ╚████║██║  ██╗███████║
   ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝   ╚═╝        ╚═╝   ╚═╝  ╚═╝╚═╝╚═╝  ╚═══╝╚═╝  ╚═╝╚══════╝

███╗   ██╗ ██████╗ ████████╗      ██╗██╗   ██╗███████╗████████╗
████╗  ██║██╔═══██╗╚══██╔══╝      ██║██║   ██║██╔════╝╚══██╔══╝
██╔██╗ ██║██║   ██║   ██║         ██║██║   ██║███████╗   ██║
██║╚██╗██║██║   ██║   ██║    ██   ██║██║   ██║╚════██║   ██║
██║ ╚████║╚██████╔╝   ██║    ╚█████╔╝╚██████╔╝███████║   ██║
╚═╝  ╚═══╝ ╚═════╝    ╚═╝     ╚════╝  ╚═════╝ ╚══════╝   ╚═╝
███████╗████████╗ ██████╗ ██████╗ ███████╗███████╗██╗
██╔════╝╚══██╔══╝██╔═══██╗██╔══██╗██╔════╝██╔════╝██║
███████╗   ██║   ██║   ██║██████╔╝█████╗  ███████╗██║
╚════██║   ██║   ██║   ██║██╔══██╗██╔══╝  ╚════██║╚═╝
███████║   ██║   ╚██████╔╝██║  ██║███████╗███████║██╗
╚══════╝   ╚═╝    ╚═════╝ ╚═╝  ╚═╝╚══════╝╚══════╝╚═╝

Memory
That Thinks.
Not Just Stored.

Your AI forgets everything between sessions. TrueMemory remembers it all, runs on one SQLite file, and never touches the cloud.

View on GitHub → See Benchmarks

One-line install

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

[02]

Not Just Cheaper.
Better at Everything.

TrueMemory doesn't just save you money. It outperforms every alternative on accuracy, features, and privacy.

TrueMemory Pro

Mem0

+31.6pp more accurate

LoCoMo Accuracy	93.0%	61.4%
Multi-hop Reasoning	96.5%	37.7%
Runs Offline	✓	✗
Personality Profiles	✓	✗
Temporal Reasoning	✓	✗
Contradiction Detection	✓	✗
Open Source	✓	✗
Cost per Answer	$0.0015	$0.0031

TrueMemory Pro

Supermemory

+27.6pp more accurate

LoCoMo Accuracy	93.0%	65.4%
Multi-hop Reasoning	96.5%	✗
Runs Offline	✓	✗
Personality Profiles	✓	✗
Temporal Reasoning	✓	✗
Contradiction Detection	✓	✗
Open Source	✓	✗
Monthly Cost	$0	$19–$399

TrueMemory Pro

RAG (ChromaDB)

+6.8pp more accurate + 6 extra capabilities

LoCoMo Accuracy	93.0%	86.2%
Multi-hop Reasoning	96.5%	✗
Personality Profiles	✓	✗
Temporal Reasoning	✓	✗
Contradiction Detection	✓	✗
Salience Filtering	✓	✗
Consolidation	✓	✗
Predictive Coding	✓	✗

[03]

Six Layers.
One Chip.

Each layer is a functional component of a custom memory processor. Together they form a single orchestration engine with graceful degradation.

L0 Personality Processor Core

Char-n-gram style vectors + entity profiles. Answers "what kind of person is X?" questions that keyword search can't touch.

~2msStyle vectorsEntity profiles

L2 Episodic Memory Registers

FTS5 full-text keyword search with temporal filtering. Fast, broad recall across the full memory store.

5-8msSQLite FTS5BM25 ranking

L3 Semantic Tensor Core Array

Dense vector search (Model2Vec or Qwen3 by tier) + RRF fusion with episodic + cross-encoder reranking. The heavy lifter.

18-22msQwen3-Embedding 0.6BRRF k=60

L4 Salience Signal Filter

Learned 13-feature logistic regression scorer trained on retrieval-utility labels. Noise filtering + entity boosting.

8-12ms13 featuresLogistic regression

L5 Consolidation Bus Controller

Structured fact summaries, contradiction resolution, and surprise-weighted reranking. Memory that evolves, not accumulates.

4-6msContradiction detectionSurprise rerank

+ Reranker Output Stage

Cross-encoder reranking (MiniLM or gte-modernbert by tier) for final precision. Orders the output by true relevance.

15-25msgte-reranker-modernbert149M params

55ms 6-layer search

468ms agentic + reranker

~1.5s agentic + HyDE

[04]

We Don't Think Benchmarks
Tell the Full Story.
But We Lead Them Anyway.

LoCoMo (1,540 questions), LongMemEval (500 questions), BEAM-1M (700 questions), BEAM-10M (200 questions). All systems share the same answer model, judge, and scoring pipeline.

93.0%

LoCoMo Accuracy

Pro Tier

+31.6pp

vs Mem0

93.0% vs 61.4%

Cost Efficiency

$0.0015 / correct answer

#	System	Accuracy	Offline	Multi-hop	Temporal
1	TrueMemory Pro	93.0%	✓	96.5%	84.4%
2	TrueMemory Base	92.0%	✓	✓	✓
3	TrueMemory Edge	89.6%	✓	✓	✓
4	RAG (ChromaDB)	86.2%	✓	✗	✗
5	Engram	84.5%	✗	✗	✗
6	BM25 Baseline	80.5%	✓	✗	✗
7	Supermemory	65.4%	✗	✗	✗
8	Mem0	61.4%	✗	37.7%	✗

[04b]

Cross-Benchmark
Validation.

Pro tier, 3-run validated means. Same judge, same scoring pipeline.

92.0%

LongMemEval Oracle

460/500 · multi-session

87.8%

LongMemEval Strict

439/500 · no partial credit

76.6%

BEAM-1M · SOTA

700 q · 35 convos · 1M+ tok

65.0%

BEAM-10M

200 q · 10 convos · 10M tok

Validate Our Results

LoCoMo Scripts → LongMemEval Scripts → BEAM Scripts → Score Verifier →

[04c]

Dive Deeper.
Read the Research.

arXiv:2605.04897 · cs.CL

Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall

Sauron Labs

[05]

Setup in 5 Minutes.
No GPU Required.

Works with Claude Desktop, Claude Code, or any MCP-compatible client. Base tier runs on any machine with 512MB RAM.

terminal

# install — Mac/Linux one-liner
$ curl -LsSf https://raw.githubusercontent.com/buildingjoshbetter/TrueMemory/main/install.sh | sh

# restart Claude, then type:
"Set up TrueMemory"

# pick Edge, Base, or Pro — that's it.
# TrueMemory captures conversations automatically from here.
# when you quit a Claude session, memories are processed
# and stored — like dreaming.

# or use the Python API directly:
from truememory import MemoryEngine
engine = MemoryEngine("memories.db")
engine.store("loves pineapple on pizza, hates cilantro")
results = engine.search("what foods does this person like?")
# → "loves pineapple on pizza, hates cilantro" (score: 0.94)

Edge Tier

Model2Vec potion-base-8M (8M params, 256d) + MiniLM-L-6-v2 reranker. Any CPU. ~30MB footprint. 89.6% LoCoMo accuracy.

Base Tier

Qwen3-Embedding-0.6B (600M params, 256d) + gte-reranker-modernbert (149M). 4GB+ RAM. ~1.5GB footprint. 92.0% LoCoMo accuracy.

Pro Tier

Same as Base + HyDE query expansion (requires LLM API key). 4GB+ RAM. 93.0% LoCoMo accuracy.

Memory
That Thinks.
Not Just Stored.

See How Much
You'll Save.

Not Just Cheaper.
Better at Everything.

Six Layers.
One Chip.

We Don't Think Benchmarks
Tell the Full Story.
But We Lead Them Anyway.

Cross-Benchmark
Validation.

Dive Deeper.
Read the Research.

Setup in 5 Minutes.
No GPU Required.

Your Agent Needs
Its Memory.

MemoryThat Thinks.Not Just Stored.

See How MuchYou'll Save.

Not Just Cheaper.Better at Everything.

Six Layers.One Chip.

We Don't Think BenchmarksTell the Full Story.But We Lead Them Anyway.

Cross-BenchmarkValidation.

Dive Deeper.Read the Research.

Setup in 5 Minutes.No GPU Required.

Your Agent NeedsIts Memory.

Memory
That Thinks.
Not Just Stored.

See How Much
You'll Save.

Not Just Cheaper.
Better at Everything.

Six Layers.
One Chip.

We Don't Think Benchmarks
Tell the Full Story.
But We Lead Them Anyway.

Cross-Benchmark
Validation.

Dive Deeper.
Read the Research.

Setup in 5 Minutes.
No GPU Required.

Your Agent Needs
Its Memory.