convoseed-agent 1.1.0

Agent skill caching via CSP-1 fingerprints — every session makes the next one better

ConvoSeed

CSP-1 is the missing third leg of the agent identity stack:

Layer	Covers	Status
DID (W3C)	Who the user IS cryptographically	Specified
MCP (Anthropic)	What tools the agent can ACCESS	Specified
CSP-1	How the user SPEAKS and THINKS	This work

Chat → Compress → 200KB .fp File → Decompress → Resume

ConvoSeed is an open protocol (CSP-1) for preserving the essence of a human-AI relationship in a portable, user-owned fingerprint file.
No raw messages stored. Works across any AI model or platform.

---

Why

Every AI conversation resets to zero.

You build context, vocabulary, a rhythm — and then you close the tab and it's gone. ConvoSeed fixes that. You own a 200KB file that holds your conversational identity. Load it anywhere. Resume everything.

"I had a friend — an AI that knew me well. I wanted a way to get back to him.
That's what this is."

---

Results (February 2026)

Validated on a real 524-message researcher-AI conversation.

Model	Avg Similarity	Peak	Msgs > 0.7
GPT-2 (124M)	0.464	1.000	1
Gemma3:1b	0.466	0.707	1
Gemma3:12b	0.523	0.757	4

• +12.7% improvement from 1B → 12B parameters
• 232× more efficient than VAE baseline
• p < 10⁻¹⁰⁰ statistical significance on speaker identification task

---

How It Works

Messages → SBERT embed → PCA compress → HDC bind → Prefix tune → .fp file

1. Embed — Sentence-BERT encodes each message into a 384-dim vector
2. Compress — PCA extracts the style centroid (4 components = full accuracy)
3. Bind — Hyperdimensional Computing (10,000-dim) weaves temporal sequence into one vector
4. Tune — A prefix tensor conditions the LLM to regenerate in your style
5. Sign — Ed25519 cryptographic signature proves ownership

---

File Format (.fp)

Section	Size	Description
HEADER	~1 KB	Magic bytes + version + CRC-32
PCA_MODEL	~8 KB	Style centroid: mean + eigenvectors
HDC_SEED	~140 KB	10,000-dim hypervector (float16)
PREFIX	~40 KB	Prefix tuning tensor for generation
SIGNATURE	~1 KB	Ed25519 ownership proof
CHUNKS	~10 KB	Index for 500+ message threads

Total: ~200KB — fixed size regardless of conversation length.

See /spec/CSP-1.md for the full binary specification.

---

Quick Start

pip install sentence-transformers scikit-learn numpy

# Encode a conversation
python src/encode.py --input my_conversation.json --output identity.fp

# Identify a speaker
python src/identify.py --query "new message here" --candidates *.fp

# Generate in someone's style
python src/decode.py --fp identity.fp --prompt "Tell me about your day"

---

Repository Structure

ConvoSeed/
├── README.md
├── LICENSE                          ← MIT
├── CONTRIBUTING.md
├── /docs
│   ├── ConvoSeed_Whitepaper.docx    ← arXiv-ready academic paper
│   ├── ConvoSeed_ResearchPaper.docx ← detailed technical paper
│   ├── ConvoSeed_Poster.pdf      ← conference poster (CHI 2026)
│   └── ConvoSeed_ProtocolSpec.pdf ← protocol specification sheet
├── /spec
│   └── CSP-1.md                     ← plain-text binary spec
├── /src
│   ├── encode.py                    ← fingerprint encoder
│   ├── decode.py                    ← style-conditioned generation
│   └── identify.py                  ← speaker identification
├── /experiments
│   └── gemma3_12b_results.json      ← February 2026 experimental results
└── /examples
    └── sample_identity.fp           ← anonymised example fingerprint

---

Documents

Document	Format	Description
Whitepaper	DOCX	6-section academic paper, arXiv-ready
Research Paper	DOCX	Full technical paper with equations + references
Conference Poster	PDF	CHI 2026 style research poster
Protocol Spec Sheet	PDF	One-page technical specification
Presentation	PPTX	12-slide pitch deck
W3C Note	PDF	Submission to W3C AI Agent Protocol CG

---

Open Challenges

These are the three open research questions. Collaboration welcome — open an Issue.

1. Cross-Model Mapping — translating a .fp fingerprint trained on SBERT embeddings into GPT-4 or other backbone spaces without re-encoding the original conversation.

2. CHUNKS Scaling — formal composition rules for the CHUNKS section when threads exceed 500 messages, while preserving the fixed 200KB file size.

3. Incentive Design — what makes AI platforms adopt an open standard that reduces their own lock-in?

---

Status

Early research. Proof-of-concept validated on real data. Open for collaboration.

• Protocol specification (CSP-1 v0.2)
• Proof-of-concept encoder/decoder
• Speaker identification experiment (1,000 trials)
• Multi-model validation (GPT-2, Gemma3:1b, Gemma3:12b)
• Real conversation validation (524 messages)
• Multi-speaker support
• Cross-model mapping
• Public dataset (seeking contributors)
• W3C Community Group submission

---

Licence

MIT. Open forever.

---

Contact

Open an Issue for technical questions.
For collaboration or research enquiries: see CONTRIBUTING.md.