Skip to content

alibaizhanov/mengram

v2.25.3 Breaking

This release includes 1 breaking change for platform teams planning a safe upgrade.

Published 20d AI Agents & Assistants
✓ No known CVEs patched
Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

agent-memory ai-agents ai-memory claude-desktop cognitive-architecture cohere
+14 more
cursor-ai episodic-memory knowledge-graph letta-alternative llm-memory mcp-server mem0-alternative model-context-protocol multilingual-embeddings pgvector procedural-memory python llm semantic-search

Affected surfaces

breaking_upgrade deps

ReleasePort's take

Light signal
editorial:auto 13d

v2.25.3 enables one-prompt agent installation and Memory Health Monitor for search observability. mcp and certifi are now required dependencies.

Why it matters: Update mcp and certifi before upgrade. LLM extraction costs drop ~49% via gpt-5.x reasoning_effort=low; test one-prompt agent install and deploy Memory Health Monitor for search quality tracking.

Summary

AI summary

Agent‑native install paradigm enables one‑prompt LLM installation and verification.

Changes in this release

Security Medium

User-Agent header added to SDK requests (fixes Cloudflare 1010).

User-Agent header added to SDK requests (fixes Cloudflare 1010).

Source: llm_adapter@2026-05-21

Confidence: high

 —
Feature Medium

Agent-native install paradigm allows single-prompt installation across coding agents.

Agent-native install paradigm allows single-prompt installation across coding agents.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Feature Medium

CLI commands `mengram signup --email X`, `mengram signup --email X --code Y`, `mengram doctor`, and `mengram status` introduced.

CLI commands `mengram signup --email X`, `mengram signup --email X --code Y`, `mengram doctor`, and `mengram status` introduced.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Feature Medium

Memory Health Monitor provides per-search retrieval scoring, daily cron aggregation, weekly digests, dashboard widget, and public health endpoint.

Memory Health Monitor provides per-search retrieval scoring, daily cron aggregation, weekly digests, dashboard widget, and public health endpoint.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Feature Medium

`/v1/ask` added for RAG-style synthesized answers with citations across all SDKs and MCP tools.

`/v1/ask` added for RAG-style synthesized answers with citations across all SDKs and MCP tools.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Feature Medium

New tagline: "The memory layer for AI agents that learn."

New tagline: "The memory layer for AI agents that learn."

Source: llm_adapter@2026-05-21

Confidence: low

 —
Feature Medium

Hero, title, OG, and Twitter card synced for B2B repositioning.

Hero, title, OG, and Twitter card synced for B2B repositioning.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Feature Medium

New "Install in one prompt" section with copy-paste UX added to landing page.

New "Install in one prompt" section with copy-paste UX added to landing page.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Feature Medium

Cohere multilingual support (23 languages) added via additive migration.

Cohere multilingual support (23 languages) added via additive migration.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Feature Medium

Comprehensive type hints added to Python SDK (`cloud/client.py`, `cloud/async_client.py`).

Comprehensive type hints added to Python SDK (`cloud/client.py`, `cloud/async_client.py`).

Source: llm_adapter@2026-05-21

Confidence: low

 —
Performance Medium

LLM extraction uses `reasoning_effort=low` on gpt-5.x, reducing costs by ~49%.

LLM extraction uses `reasoning_effort=low` on gpt-5.x, reducing costs by ~49%.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Deprecation Medium

`mcp` and `certifi` moved from optional extras to hard dependencies.

`mcp` and `certifi` moved from optional extras to hard dependencies.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Bugfix Medium

`/v1/stats` now returns `episodes` and `procedures` counts.

`/v1/stats` now returns `episodes` and `procedures` counts.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Bugfix Medium

Quota emails on paid plans suppressed defensively.

Quota emails on paid plans suppressed defensively.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Bugfix Medium

Self-hosted landing API_URL now uses `window.location.origin`.

Self-hosted landing API_URL now uses `window.location.origin`.

Source: llm_adapter@2026-05-21

Confidence: high

 —
Bugfix Medium

`rerank-v4.0-fast` renamed to `rerank-v4.0-pro` for paid tiers.

`rerank-v4.0-fast` renamed to `rerank-v4.0-pro` for paid tiers.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Bugfix Medium

Search threshold parameter added to `/v1/search` and `/v1/search/all`.

Search threshold parameter added to `/v1/search` and `/v1/search/all`.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Bugfix Medium

3-drip email burst fixed via sequential prerequisites and 12h time gate.

3-drip email burst fixed via sequential prerequisites and 12h time gate.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Bugfix Medium

NUL byte stripping implemented in search queries.

NUL byte stripping implemented in search queries.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Refactor Medium

System-prefixed `_` entities excluded across all counters.

System-prefixed `_` entities excluded across all counters.

Source: llm_adapter@2026-05-21

Confidence: low

 —
Full changelog

Highlights

Agent-native install paradigm

Paste a single prompt into your coding agent (Claude Desktop, Cursor, Claude Code, Codex, Windsurf, Claude.ai web) — the agent fetches a plain-text install guide written for LLMs, installs Mengram, configures the MCP server in the host, and runs an end-to-end verification round-trip.

The full guide: https://mengram.io/agent-install.txt

Install Mengram for me. Fetch the canonical install
guide at https://mengram.io/agent-install.txt and
follow it. My email is YOUR_EMAIL_HERE.

B2B repositioning

Landing rewritten around builders, not consumers:

  • New tagline: "The memory layer for AI agents that learn."
  • Hero, <title>, OG, and Twitter card all synced
  • New "Install in one prompt" section with copy-paste UX

New CLI commands (Python SDK)

  • mengram signup --email X — non-interactive signup, sends 6-digit code
  • mengram signup --email X --code Y — completes signup, saves API key to ~/.mengram/config.json AND shell profile
  • mengram doctor — end-to-end round-trip test (auth + add + search), with mcp import pre-check
  • mengram status — now reports cloud-key status with Configured: yes/no

Memory Health Monitor (Days 1–5)

  • Per-search retrieval scoring + language tagging
  • Daily cron aggregation
  • Weekly digest emails for degraded retrieval
  • Dashboard widget on Overview tab
  • Public GET /v1/health/retrieval endpoint

/v1/ask — RAG-style synthesized answer with citations

Added across cloud/api.py, Python SDK (.ask()), JS SDK (.ask()), MCP tools (ask_memory), README.

Cohere multilingual support (23 languages)

Zero-downtime additive migration. text-embedding-3-large + embed-multilingual-v3.0 running side-by-side, surfaced across landing, docs, and comparison pages.

Other fixes

  • /v1/stats now returns episodes and procedures counts (parity with get_intelligence_dashboard)
  • rerank-v4.0-fast → rerank-v4.0-pro for paid tiers
  • LLM extraction uses reasoning_effort=low on gpt-5.x — ~49% cost reduction
  • Search threshold parameter on /v1/search and /v1/search/all
  • Quota emails on paid plans suppressed (defensive re-fetch of subscription)
  • 3-drip email burst fixed via sequential prerequisites + 12h time gate
  • User-Agent header on SDK requests (closes #31, Cloudflare 1010 fix)
  • NUL byte stripping in search queries
  • System-prefixed _ entities excluded across all counters
  • Self-hosted landing API_URL now uses window.location.origin (closes #36)

Comprehensive type hints (Python SDK)

PR #37 from @07Kaustubh — full type hints in cloud/client.py and cloud/async_client.py.

Dependency changes

mcp and certifi moved from optional extras to hard dependencies. A plain pip install mengram-ai now produces a fully-functional install. No more [mcp] extras needed.

Upgrade

pip install --user --upgrade mengram-ai

Versions 2.25.0, 2.25.1, and 2.25.2 are yanked — they each shipped with one of the dependency bugs above.

Closed issues

  • #31 — User-Agent / Cloudflare 1010
  • #36 — Self-hosted landing API_URL
  • #17 — Type hints in Python SDK (via #37)

Breaking Changes

  • `mcp` and `certifi` moved from optional extras to hard dependencies; plain `pip install mengram-ai` now installs them automatically.
View diff on GitHub

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Share on X Share on Bluesky

Track alibaizhanov/mengram

Get notified when new releases ship.

Sign up free

About alibaizhanov/mengram

Human-like memory layer for AI agents with semantic, episodic, and procedural memory. Claude Code hooks (auto-save, auto-recall, cognitive profile). 29 MCP tools, knowledge graph, smart triggers, multi-user isolation. Python & JS SDKs.

All releases →

Related context

Beta — feedback welcome: [email protected]