convoseed-agent

Compress any conversation into a portable 200KB .fp fingerprint file.

ConvoSeed implements the CSP-1 protocol — a method for encoding the style of a conversation (not the content) using SBERT embeddings, PCA compression, and Hyperdimensional Computing. The result is a fixed-size file you own and can load into any AI session.

pip install convoseed-agent

5-minute demo

Step 1 — install

pip install convoseed-agent
pip install sentence-transformers scikit-learn numpy

Step 2 — encode a conversation

Your conversation must be a JSON file in this format:

[
    {"role": "user", "content": "I've been thinking about memory..."},
    {"role": "assistant", "content": "Memory is deeply selective..."},
    ...
]

Then run:

convoseed-encode --input my_conversation.json --output identity.fp

Or in Python:

import json
from convoseed_agent import encode_conversation

with open("my_conversation.json") as f:
    messages = json.load(f)

encode_conversation(messages, "identity.fp")
# → identity.fp  (~200KB, fixed size regardless of conversation length)

Step 3 — identify a speaker from a new message

from convoseed_agent import identify, load_fp
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

winner, scores = identify(
    query_text="I wonder if that cognitive style could be captured somehow",
    fp_paths=["identity.fp", "other_person.fp"],
    model=model
)

print(f"Best match: {winner}")
for path, score in sorted(scores.items(), key=lambda x: -x[1]):
    print(f"  {score:.4f}  {path}")

What it does

Messages → SBERT embed → PCA compress → HDC bind → .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. Save — Written to a portable JSON-based .fp file (~200KB)

Key result from the research paper: 4 PCA components capture full speaker identification accuracy, meaning conversational style is genuinely low-dimensional. You can represent how someone thinks with 4 numbers.

Research results

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

Speaker identification: 52% accuracy on 10 candidates (vs 10% random baseline), p < 10⁻¹⁰⁰.

Optional: generation (requires torch)

pip install convoseed-agent[decode]

from convoseed_agent import generate_with_prefix, load_fp

fp = load_fp("identity.fp")
output = generate_with_prefix("Tell me about your weekend", fp, model_name="gpt2")
print(output)

Status

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

• CSP-1 protocol specification
• Encoder / decoder / identifier
• Speaker identification experiment (1,000 trials)
• Multi-model validation
• Cross-model mapping (open research problem)
• Public fingerprint registry

Links

• GitHub: https://github.com/0xAshraFF/ConvoSeed
• Protocol spec: /spec/CSP-1.md in the repo
• Research paper: /docs/ in the repo

MIT License.