Skip to main content

Self-hostable framework for developing coding-agent context from lived experience

Project description

Dreamer - self-evolving context for your coding agents

Image

Get started · Feedback and pruning · Extensions · Blogpost

Dreamer keeps your team's AGENTS.md and skills up to date with what your coding agents learn while they work. It runs as a self-hostable MCP server that collects memories from every agent on the team and, on a schedule, regenerates the context bundle the next session reads.

Team-wide memory. Memories from every agent on the team pool into a single store and feed a single context bundle, instead of staying on one workstation.

Any coding CLI. Anything that speaks MCP submits memories through the same submit_memory tool, including Claude Code, Cursor, Codex, and custom agents.

Extendible by config. STM store, LTM store, context store, dream engine, auth, triggers, and hooks are Python Protocols wired up from YAML. Swap any default by pointing at a different class.

Memory that prunes itself. Agents report whether context guidance held up in practice. Dreams reinforce confirmed entries and retire contradicted ones, archiving rather than deleting.


Dreamer is part of the broader LUML effort to build open infrastructure for autonomous ML agents.

Core — registry, deployments, monitoring Prisma — autonomous ML research agents Flow — experiment tracking and tracing

Get started

Dreamer requires Python 3.12 or later. The defaults extra pulls in SQLite for STM, the Claude Agent SDK for dreaming, APScheduler for cron triggers, and gitpython for the post-dream commit hook.

pip install 'dreamer-server[defaults]'

Scaffold a project.

dreamer init

This writes a dreamer.yaml, a workspace/ with memory/ and context/ subdirectories, and a .gitignore that keeps the SQLite database out of git.

Issue a token for your agents to send in the Authorization header.

dreamer-simple-auth token create --db ./dreamer.db --name my-token

Sanity-check the config. The loader resolves every component, runs the protocol-conformance check, and prints the wired graph and per-slot multi-tenancy table.

dreamer config check

Run the server.

dreamer serve

Point Claude Code or any MCP client over streamable-http at http://localhost:8080/mcp/ with Authorization: Bearer <token>. The server advertises a submit_memory tool whose accepted types come from your config — out of the box observation, failure, and code_snippet — plus the confirm_context and flag_context feedback tools described below.

Cron is the default trigger. To fire a one-shot dream from the command line:

dreamer dream --trigger external

Feedback and pruning

Generated context marks guidance derived from long-term memory with a visible [mem: <slug>] marker and instructs agents to report back through two MCP tools. confirm_context records that the guidance proved useful. flag_context records that it proved wrong, with the observed evidence. Flags may name target slugs, quote the misleading text, or carry neither.

Feedback goes through the same pipeline as submit_memory, but the dream receives confirmations as a single aggregate report, and the scaffolded threshold trigger excludes feedback types from its count. Feedback alone does not trigger a dream.

Each dream applies a conservative policy. Evidenced flags outweigh missing confirmations, and age is only a tiebreaker. A single flag weakens a well-confirmed entry; corroborated flags supersede it. Retiring an entry moves it to an archive/ subtree with retirement frontmatter, and every decision is logged in archive/LOG.md.

The store enforces safety rails at commit time:

ltm_store:
  class: dreamer.contrib.ltm.markdown.MarkdownLTMStore
  params:
    root: ./workspace/memory
    max_autonomous_removals: 5   # archival moves are exempt
    enforce_pinned: true         # `importance: pinned` entries stay put
    on_guard_violation: fail     # or `warn`

A violating commit fails before any mutation and the batch returns to STM for the next run.

Extensions

Dreamer is config-assembled. dreamer.yaml wires module.path.ClassName references into a component graph. Every slot sits behind a Python Protocol defined in dreamer.api, including the STM store, the LTM store, the context store, the dream engine, auth, triggers, and hooks. The shipped defaults are chosen to get a team running in a few minutes, and every one of them can be swapped.

stm_store:
  class: dreamer.contrib.stm.sqlite.SQLiteSTMStore
  params:
    path: ./data/stm.db

ltm_store:
  class: dreamer.contrib.ltm.markdown.MarkdownLTMStore
  params:
    root: ./workspace/memory

context_store:
  class: dreamer.contrib.context.markdown.MarkdownContextStore
  params:
    root: ./workspace/context

dream_engine:
  class: dreamer.contrib.dream.claude_agent.ClaudeAgentDreamEngine

triggers:
  - class: dreamer.contrib.triggers.cron.CronTrigger
    params:
      schedule: "0 */6 * * *"

To plug in your own component, write a class that satisfies the protocol. For example, a Postgres-backed STM store:

from typing import ClassVar
from dreamer.api.compat import implements
from dreamer.api.stores import STMStore

@implements(STMStore, version=1)
class PostgresSTMStore:
    multi_tenant: ClassVar[bool] = True

    def __init__(self, *, dsn: str) -> None:
        ...

    async def submit(self, memory, *, ctx): ...
    async def claim_batch(self, *, ctx): ...

Then reference it from dreamer.yaml:

stm_store:
  class: my_pkg.stores.PostgresSTMStore
  params:
    dsn: ${env:POSTGRES_DSN}

dreamer config check validates the protocol version, signatures, parameter kinds, and capability requirements before the server boots. The same shape of change covers a graph-backed long-term memory store, an OIDC auth backend, or a Slack notification hook in place of the git commit.

dreamer.testing.conformance ships abstract pytest classes for each protocol. The cases cover idempotency, lease isolation, expired-lease reclamation, tenant-scope leakage, and the purge_consumed contract. Any compliant implementation should pass them.

from dreamer.testing.conformance.stm_store import STMStoreConformance

class TestPostgresSTMStore(STMStoreConformance):
    @pytest.fixture
    async def store(self):
        return PostgresSTMStore(dsn="postgresql://...")

License

MIT. See LICENSE.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

dreamer_server-0.2.0.tar.gz (247.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

dreamer_server-0.2.0-py3-none-any.whl (152.7 kB view details)

Uploaded Python 3

File details

Details for the file dreamer_server-0.2.0.tar.gz.

File metadata

  • Download URL: dreamer_server-0.2.0.tar.gz
  • Upload date:
  • Size: 247.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for dreamer_server-0.2.0.tar.gz
Algorithm Hash digest
SHA256 ba001769c01ad21bd7c567b8112fba2b082a101672fa7e8a5e7b7e4ac69ab238
MD5 850d40e7a2e70d5819eb58748a5f3943
BLAKE2b-256 9bb8d95912ef3c378cf8fe7238382ccdb2231748c572d14c587138c496f10a4c

See more details on using hashes here.

Provenance

The following attestation bundles were made for dreamer_server-0.2.0.tar.gz:

Publisher: publish.yml on luml-ai/dreamer

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file dreamer_server-0.2.0-py3-none-any.whl.

File metadata

  • Download URL: dreamer_server-0.2.0-py3-none-any.whl
  • Upload date:
  • Size: 152.7 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for dreamer_server-0.2.0-py3-none-any.whl
Algorithm Hash digest
SHA256 727005446b0e6bf06ec11a0581c5c7898862c1239f067d767505c0fa50078b69
MD5 b6ccfed64323811efd8373a8f4eaf1eb
BLAKE2b-256 8042ec2e951a9645505cbdbcf00e4ed695b77f6a2b862e9744acfd29ea881106

See more details on using hashes here.

Provenance

The following attestation bundles were made for dreamer_server-0.2.0-py3-none-any.whl:

Publisher: publish.yml on luml-ai/dreamer

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page