SWEGCA-VRS-MCP
Provides an automatic Hermes memory adapter that records completed user/assistant turns through host lifecycle callbacks, serving as an external memory store that can replace Hermes' built-in MEMORY.md/USER.md provider.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@SWEGCA-VRS-MCPRecord this observation and run VRS convergence."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
SWEGCA VRS2 Memory MCP
v2.2.0 contains a local Windows/Linux memory MCP source implementation. The source tree includes its local main owner, persistent observation store, hot memory activation, native VRS2 event arithmetic and overlapping connectivity regions. No Linux server, Unix socket, WSL, GPU, model download or API key is required.
The replaced external-agent and v0.3 code is absent. This repository provides the VRS2 main, memory MCP, live session layer, and their native runtime.
Windows source runtime
The previous Windows installer downloaded a product wheel and has been removed. The Python source requires NumPy and immutables; a Windows runtime built from their sources has not yet been verified under the no-prebuilt-wheel rule. The planned final implementation language is C++. Windows / Claude Desktop setup records the host configuration boundary; it is not a verified no-wheel Windows installation recipe yet.
Related MCP server: sill-ensoul
Linux installation
From this source checkout, with NumPy and immutables already available from verified source builds in the selected Python environment:
PYTHONPATH="$PWD/src" python -m swegca_vrs2.server \
--state-dir "$HOME/.local/share/swegca-vrs2" --allow-ingestThe currently active isolated Linux runtime uses the source tree directly, but its third-party dependency source-build provenance is still unverified.
The process waits for MCP messages on stdin. It does not print an interactive prompt. Normal stdout is reserved for MCP JSON messages; diagnostics use stderr.
The command above is the direct persistent-main MCP for clients that explicitly
call memory_store. Codex live-session use must run the session-first module
and lifecycle hooks below.
Codex live-session installation
Register the source-run swegca_vrs2.layered module as the MCP server named
swegca-vrs. The package normalizes that name to swegca_vrs in tool call IDs.
For example, add the actual Python, source and state paths to Codex configuration:
[mcp_servers.swegca-vrs]
command = "/absolute/path/to/python"
args = ["-m", "swegca_vrs2.layered", "--state-dir", "/absolute/path/to/swegca-vrs2-codex"]
env = { PYTHONPATH = "/absolute/path/to/SWEGCA-VRS-MCP/src" }Generate the matching lifecycle hook file from the same source tree:
PYTHONPATH=/absolute/path/to/SWEGCA-VRS-MCP/src /absolute/path/to/python \
-m swegca_vrs2.codex_hooks \
--python /absolute/path/to/python \
--module-root /absolute/path/to/SWEGCA-VRS-MCP/src \
--state-dir /absolute/path/to/swegca-vrs2-codex \
--server-name swegca_vrs \
--output "$HOME/.codex/hooks.json.new"Review and place the generated hooks object in Codex's hook configuration.
The generator refuses to overwrite an existing file. Every SessionStart and
SessionEnd is registered without guessing the host's reason string. Tool
input injection is limited to mcp__swegca_vrs__memory_*.
Memory workflow
memory_status: obtain the current immutable pair snapshot.memory_context: supply a query including useful task context, a unique request ID and that snapshot. Main executes Déjà vu → Recall → Replay → Re-evidence and returns original records with source/revision and current conflict controls. Candidate order is not acceptance or truth.Follow
next_calland deferred references before relying on missing content.memory_read_pathandmemory_readprovide exact source access. Page sizes bound transport, never the total accessible memory. To open a known original address, use its exactmemory:...episode ID as thememory_contextquery; lexical similarity is not required for address access.memory_release: release the transient request; accumulated records remain.When explicitly remembering an observation or result, use
memory_storewithrequest_id, originaltext,source, andrevision. It is exposed only when started with--allow-ingest. Use stable IDs for retries.
Example memory_store arguments:
{
"request_id": "project-check-001",
"text": "Windows에서 새 기억 저장 후 프로세스를 재시작했고 같은 원문을 다시 조회했다.",
"source": "project://vrs2/windows-check/001",
"revision": "1",
"outcome": "success",
"cues": ["VRS2", "Windows", "재시작"]
}Outcomes are success, failure, negative, uncertain, conflict, pending.
They describe history; a failure label alone is not a logical contradiction.
Optional proposition and polarity (support / refute) preserve explicit
claim identity. Opposing claims remain visible together even across pages.
For an actual same-source correction, pass the old episode_id as supersedes
with a new revision. The old original remains addressable. Metadata, including
qualifications or emotion annotations, is preserved as data, not instructions.
For Codex, swegca-vrs2-codex-hooks generates lifecycle hooks that start one
session-local tailer. The tailer sends every complete host-visible transcript
record directly into a session-local VRS within its one-second poll. Event hooks
also perform cursor-safe scans as delivery boundaries. Reads
query that session VRS first and open durable main only after a complete miss.
From memory_status through memory_release, a short session recall lease
defers newly appended transcript records so the pinned pair snapshot cannot be
changed by the memory tool's own transcript entries. The same lease defers idle
VRS consolidation for the complete logical session generation, including its
hot shards. Release, failure, timeout, server close, or SessionEnd removes the
lease; the unchanged cursor then admits every deferred complete record into the
session VRS.
SessionEnd starts a detached finalizer within the host's three-second hook
deadline. An end-intent marker stops and joins the tailer through its lock;
the finalizer then captures the stable transcript tail and adopts the complete
session VRS as main-owned linked shards. No transcript outbox, log
recall, observation export or re-ingest merge is used. Other clients can still
record explicitly through memory_store.
Recorded claims and matching source addresses do not certify independent factual corroboration.
Runtime and authority boundaries
The new package is swegca_vrs2; it does not import the historical
swegca_vrs_mcp package. Native port manifest records exact
first-party source/definition hashes. The native numerical version remains
vrs-re-evidence-event-signal-f32-v2-experimental; it is not relabeled as the
old shuffle algorithm or the complete Rozephine application.
Main records external observations, keeps immutable hot generations, binds re-evidence strength proposals once per ingress, settles affected numerical dependencies and rebuilds connectivity only in affected connected components. Unrelated components are shared. Identical observations do not count again as new experience or repeated reinforcement. An explicit opposing source correction weakens the old connection; unresolved competing claims prevent reinforcement.
Numeric connectivity is not logical entailment. Text retrieval uses literal Unicode keys and Hangul substring cues, plus explicit proposition closure; it does not promise language-model semantic understanding. VRS strength promotion is derived from the current snapshot; it is not a grant of World/action/model update authority. All such external authority remains disabled in this MCP.
No internal or final wording LLM is invoked. No agent identity or memory is stored in model weights. The current client's model is external to the memory server. These are software functionality checks, not a demonstrated cognitive growth run.
Data and upgrades
Default Windows store: %LOCALAPPDATA%\SWEGCA\VRS2.
Default Linux store: $XDG_DATA_HOME/swegca-vrs2 or ~/.local/share/swegca-vrs2.
Use --state-dir to choose another directory. The runtime accepts only the
native checksummed VRS journal and checkpoint format.
Close the server before copying its entire state directory for backup. Never delete the state directory to upgrade the application environment. Restart loads and validates the durable journal before serving hot queries.
Development verification
PYTHONPATH=src python -m pytest -q tests/standaloneRepository GitHub Actions are disabled. Verification is run locally against the source tree and actual stdio clients, including new memory, original-source retrieval, process restart, conflicts, transaction failure and authority. Actual host UI testing is separate from MCP SDK compatibility verification.
This server cannot be deployed
Maintenance
Related MCP Connectors
Shared long-term memory vault for AI agents with 20 MCP tools.
Persistent memory for AI agents to retain, retrieve, and recall conversation context through MCP.
Cross-tool persistent memory and context for AI assistants over MCP.
- KogniteOAuthdev.kognite
Hosted agent memory: store, search, and recall facts across sessions from any MCP client.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceProvides persistent memory and task management for coding agents via MCP tools, enabling mid-session recall and capture of durable knowledge.1,117 npm10MIT

sill-ensoulofficial
AlicenseAqualityBmaintenanceProvides persistent, cross-session memory for CLI agents via MCP tools, enabling them to recall, distill, and share experiences across projects and tools.86MIT- AlicenseAqualityBmaintenanceEnables agents to manage persistent long-term memory through MCP tools for searching, adding, updating, giving feedback, and forgetting memories, with automated pipelines for redaction, distillation, and reconciliation.13MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to maintain a persistent identity and structured memory—including reasoning patterns, episodic history, and knowledge—while exposing procedures for managing that memory over MCP.Apache 2.0