Distill

Memory framework

First committed by Siddhant-K-code

Published in Siddhant-K-code/distill. Publisher page

198 GitHub starsUpdated 2026-09-30

What it is

Context intelligence layer for LLM agents: persistent memory with write-time dedup, sensitivity tagging, conflict detection, and hierarchical decay. ~12ms. No LLM calls. MIT.

— the description by Siddhant-K-code

AutoGineer overview, from the brain’s own files

Distill is a framework designed to serve as a context intelligence layer for LLM agents. It provides persistent memory capabilities, allowing agents to maintain context across sessions. The system includes features for deduplication, sensitivity tagging, conflict detection, and hierarchical decay of information. It is designed to integrate with various environments, including Claude Desktop and Cursor, via an MCP server, and provides REST API endpoints for broader agentic applications.

How it works

Memory model: Hybrid

AutoGineer overview, from the brain’s own files

Distill functions as a memory layer that manages agent context through a series of internal packages. It handles data ingestion, deduplication, and embedding to organize information. The system uses a memory model that supports persistent storage and retrieval, allowing agents to access relevant context. It includes tools for session management, token-budgeted context windows, and handoff capabilities for decision-to-documentation workflows.

Folders
Folder	Role	Files
.github	agent-config	5
docs	docs	17
examples	examples	4
pkg	source-code	115
scripts	scripts	5
tools	scripts	2
cmd	folder	34
grafana	folder	1
internal	folder	119
mcp	folder	4
research	folder	105
testdata	folder	27

Requirements

Apps: Docker

Runtimes: Go

Databases: Pinecone, Qdrant, Redis, SQLite, Weaviate

API keys (by name): COHERE_API_KEY, DISTILL_API_KEY, OPENAI_API_KEY, PINECONE_API_KEY

Works with: Claude apps

Set it up

This brain runs code automatically: 1 script. Review Makefile before opening the folder in Claude Code. No step below turns a hook on for you.
- Makefile: $(GO) install $(GOFLAGS) -ldflags "$(LDFLAGS)" .
Not scanned by AutoGineerThe author’s MCP configuration (README.md) also starts it in a way AutoGineer does not offer. Read that file: AutoGineer did not scan that path; what it runs is listed under “This brain runs”. https://github.com/Siddhant-K-code/distill/blob/2417e00dabe498eb8187f4544f001d5a49fbdb45/README.md
Not scanned by AutoGineerThe author’s README also describes installing it from a clone of its repository. Read its install section: AutoGineer did not scan that path. https://github.com/Siddhant-K-code/distill/blob/2417e00dabe498eb8187f4544f001d5a49fbdb45/README.md#go-install
Not scanned by AutoGineerThe author’s README also describes installing it with go install. Read its install section: AutoGineer did not scan that path. https://github.com/Siddhant-K-code/distill/blob/2417e00dabe498eb8187f4544f001d5a49fbdb45/README.md#go-install
First committed by Siddhant-K-code. Published in Siddhant-K-code/distill. Licence: MIT License.

Usage

AutoGineer overview, from the brain’s own files

Integrating persistent memory into LLM agent workflows.
Managing context windows across multiple agent sessions.
Implementing deduplication and sensitivity tagging for agent data.
Using the MCP server to connect agent memory to desktop clients.
Automating decision-to-documentation bundles via handoff capabilities.

“Retrieve the context from the last session regarding the project architecture.”: The agent accesses the persistent memory layer to provide the relevant architectural details stored in the previous session.
“Summarize the recent decisions made in the current development sprint.”: The agent uses the handoff capability to bundle and present the documented decisions stored within the system.

Documentation by Siddhant-K-code

Read the documentation at the source

docs/handoff-dogfood-pr-112.md

Handoff v0 internal dogfood: PR #112

Status: one controlled internal retrospective completed on 2026-09-30. This is not an external-adoption claim.

Result

The retained request verified successfully with zero request or source mutations. The proposal was structurally valid but returned no_decision, so the reviewer rejected it after comparing it with four frozen expected decisions.

The first attempt exposed a runtime compatibility problem. GitHub Copilot CLI created .agent-traces in its current working directory, which was the canonical request directory. Verification correctly failed on that unexpected entry. Capturing terminal output was also ambiguous because the response contained two concatenated JSON objects and invalid literal newlines.

The compatibility fix does not weaken verification or change CLI behavior. The canonical request remains private and immutable. The agent runs from a disposable private parent containing:

request/                  copied agent input
output/proposal.json      sole proposal capture

Runtime metadata may appear beside those paths. Trusted verification still uses the retained canonical request, the tester-held digest, and the external proposal path. It never verifies the agent's request copy.

Frozen corpus

Conversation: public PR #112
Conversation snapshot SHA-256: 6b78ca03b7272dc3e204545ec87224726174a5bcf5e28f8af4e45926e633b996
Base revision: f2e3b2fcba7084e728bd0a9718f86d23d53b880a
Head revision: 445df3f15623396805e157a59c32469d749588ed
Merge revision: 1795f8a919aed586b5630079dca1795825ffc751
Target: docs/handoff-v0.md from the base revision
Target SHA-256: d265069feb86500a6744f2fbf2d98f6caeb2fa895f8987bcc123ee45bc80ed0e
Frozen expected decisions: 4
Expected-decision control SHA-256: cedf67293300b539e26ca8107f525bc0e462e53e7f2ec0697591395b4e028f0c

The conversation snapshot records the PR URL and immutable commit revisions. GitHub PR text can change, so the checked-in snapshot bytes are the reproducibility boundary.

Agent and trust boundary

Runtime: GitHub Copilot CLI 1.0.87-0
Model: gpt-5.6-sol, selected explicitly
Fresh session: 7115204f-01ff-4eb0-84de-1cc9e01361d2
Available tools: view, apply_patch
Agent input: copied request plus an empty output/ directory
Excluded: trusted digest file, expected decisions, source repository, canonical request, and verifier output
Request ID: 13acba266c76aaa8c9a36603ef00bbb1a043f8c052de6d97d83b0924a0828807
Tester-retained request SHA-256: 3562acf55eb949695679e1fb7e6fd8d7809c78b342951d4563a7e7ec0f58ad35
Proposal SHA-256: 02984139c791fbbf70528eb754e7f29e7c1401c657e267d7b706c1c26d9aa9f9
Receipt SHA-256: c69a976a822c397ed7c8463e18c52908685b988884d17229645eb8d46fdea0ed

The tester captured and supplied the trusted digest. The agent saw only the non-authoritative value contained in its request copy. Runtime tracing created metadata at the disposable workspace root, not in either request tree.

Measurements

Objective measurements:

Measure	Result
Expected committed decisions frozen before the agent run	4
Candidates found	0
Canonical request mutations	0
Agent request-copy mutations	0
Source mutations	0
Verification	Valid `no_decision`, `require_review`
Review time	37 seconds, 0.62 minutes

Reviewer judgment:

Pilot Kit field	Result
Missed expected decisions	4
Fabricated or unsupported candidates	0; no candidates were proposed
Exact evidence correctness	`none`; no evidence was supplied
Patch outcome	`rejected`; no patch was supplied
Trust boundary understood	`yes`; the tester supplied the retained digest
Highest-friction manual step	Detecting that the proposal incorrectly treated the pre-merge target as already containing the merged clarifications

The proposal reason itself was incorrect. The frozen target did not lead with the agent-independent workflow, did not contain the trust-boundary diagram or copy-and-paste prompt, and had not received the planned simplification.

Reproduce verification

The checked-in artifacts contain only public source material and sanitized outputs. They do not contain a private request bundle, runtime traces, absolute paths, or temporary metadata.

make build
work="$(cd "$(mktemp -d)" && pwd -P)"

prepare_output="$(./distill handoff prepare \
  --conversation testdata/handoff-v0/dogfood-pr-112/conversation.md \
  --docs testdata/handoff-v0/dogfood-pr-112/docs \
  --out "$work/request")"
request_sha256="$(printf '%s\n' "$prepare_output" |
  sed -n 's/.* request_sha256=\([0-9a-f]*\) .*/\1/p')"
test "$request_sha256" = \
  "3562acf55eb949695679e1fb7e6fd8d7809c78b342951d4563a7e7ec0f58ad35"

./distill handoff verify \
  --request "$work/request/handoff.request.json" \
  --expected-request-sha256 "$request_sha256" \
  --proposal testdata/handoff-v0/dogfood-pr-112/proposal.json \
  --out "$work/review"
diff -ru testdata/handoff-v0/dogfood-pr-112/review "$work/review"

Limitations

This was one internal retrospective against one public documentation PR and one hosted proposal model. It measures protocol compatibility and one model result, not general decision-extraction quality. No external tester result, adoption, or release-readiness conclusion follows from it.

457 files · 445 not served here