Explicit memory / Unpublished development preview
Remember a fact.
Restart. Recall it.
This example follows one project fact through a real service shutdown and restart. The new process must return the same fact, with its source evidence, through both the CLI and MCP.
The existing benchmark release gate failed. The source and runnable example remain available for evaluation; passing this persistence check does not clear that release gate. Read the gate result.
The example uses real local embeddings and one explicit synthetic fact. All five experimental modules are disabled. It does not test automatic extraction, ranking quality, or native Hermes host acceptance.
Run it in a fresh environment
Use Python 3.12 and Git in an empty working directory. These commands
install the development source, currently versioned
0.7.0b1. They skip the large archived evaluation assets.
The first run downloads the FastEmbed model, about 100 MB. No hosted
LLM or judgment service is required.
Verified source commit:
7d5fc27757ce0a558b42677631ce943238ef789d. The checkout below selects this exact revision.
$env:GIT_LFS_SKIP_SMUDGE = "1"
git clone --depth 1 --branch fix/mnemos-delivery-20261026 https://github.com/anthony-maio/mnemos.git mnemos-preview
$mnemosCommit = "7d5fc27757ce0a558b42677631ce943238ef789d"
git -C mnemos-preview fetch --depth 1 origin $mnemosCommit
git -C mnemos-preview checkout --detach $mnemosCommit
if ((git -C mnemos-preview rev-parse HEAD) -ne $mnemosCommit) { throw "Unexpected source revision" }
python -m venv .venv
.\.venv\Scripts\python.exe -m pip install "./mnemos-preview[mcp]"
.\.venv\Scripts\python.exe mnemos-preview/examples/restart_recall.py --output restart-recall.json
macOS and Linux commands
GIT_LFS_SKIP_SMUDGE=1 git clone --depth 1 --branch fix/mnemos-delivery-20261026 https://github.com/anthony-maio/mnemos.git mnemos-preview
mnemos_commit="7d5fc27757ce0a558b42677631ce943238ef789d"
git -C mnemos-preview fetch --depth 1 origin "$mnemos_commit"
GIT_LFS_SKIP_SMUDGE=1 git -C mnemos-preview checkout --detach "$mnemos_commit"
test "$(git -C mnemos-preview rev-parse HEAD)" = "$mnemos_commit" || exit 1
python3 -m venv .venv
./.venv/bin/python -m pip install "./mnemos-preview[mcp]"
./.venv/bin/python mnemos-preview/examples/restart_recall.py --output restart-recall.json
The example creates its own temporary configuration, SQLite database, runtime directory, scope registry, and two empty Git repositories. It stops the service it started and removes the temporary workspace. Your existing memory database and host configuration are not used. The embedding model may remain in its normal download cache.
What happens inside the example
- Register two separate projects.The notebook project will own the memory. An unrelated project provides the isolation check.
-
Store one explicit fact and wait for completion.
The sample notebook project keeps release checklists in docs/release.md. - Stop the service and start a new process.The second process must retrieve the same fact ID and content from SQLite, with source evidence intact.
- Connect an MCP client and inspect the fact.The client initializes stdio, discovers the tools, recalls the fact, and reads its evidence. The unrelated project must receive an empty result.
The verification receipt
A successful run prints a JSON receipt and writes it to
restart-recall.json. The captured receipt records these
checks from a locally built, unpublished wheel in a fresh Windows
Python 3.12 environment. Compare it with your own run.
- Explicit fact committed
- Passed
- Service process restarted
- Passed
- Same fact ID recalled
- Passed
- MCP tools discovered
- 18
- MCP recall and evidence inspection
- Passed
- Unrelated project returned no facts
- Passed
- Hermes entrypoint loaded
- Passed
- Owned service stopped on exit
- Passed
Loading a Hermes entrypoint establishes package registration only. It does not establish a working native Hermes session. The MCP check uses a real protocol client, not an interactive agent host.
Use the result for the right decision
If your immediate problem is carrying an explicit repository convention across sessions, this is a useful starting point. It demonstrates a small workflow you can inspect and reproduce. The automatic enrichment pipeline has a separate question to answer: does its extra work improve what the agent retrieves, including knowing when memory contains no answer?
For this preview, begin with explicit facts and keep the research modules disabled. Review the source and evaluation before connecting a real project or enabling additional providers.