Reference · MCP server

Fourteen tools, one memory graph.

Conxt exposes your memory as a remote MCP server. Any MCP-compatible client — Claude Code, Cursor, Codex, Claude Desktop — can read the context it needs and write back what it learns, over these fourteen tools.

The framework in one rule

Open with get_context. Close with add_memory. Everything else is a sharper instrument for a narrower moment.

The failure mode isn't picking the wrong tool — it's agents that read nothing at the start and write nothing at the end, so the graph never compounds.

Connect · mcp.jsonremote

{
  "conxt": {
    "url": "https://mcp.conxt.dev/mcp/",
    "headers": { "Authorization": "Bearer cnxt_your_key" }
  }
}

memory typesdecisionpreferenceentityopen_questioncoding_ruletool_choiceworkflow

Reading

Before the agent acts — pull the context that's relevant now.

read

`get_context`

The default session-opener. Semantic search that blends personal and team memory, ranked by hybrid score (similarity + recency + reinforcement).

get_context(ctx, query, user_id=None, limit=15)

When to useCall once at the start of any task — "what do I already know that's relevant here." If your agent calls one tool per session, make it this one.

Parameter	Type	Default
`query`	string	required	What you're working on, in natural language.
`limit`	int	15 (max 30)	How many memories to return.
`user_id`	string	auto	Resolved from the API key; only pass to override.
`ctx`	Context	injected	MCP request context — supplied by the runtime, never by you.

ReturnsA formatted block of the most relevant memories, plus your pinned set. "No relevant memories found" when the graph has nothing close.

read

`resolve_context`

The verified retrieval loop. Runs discover → verify → iterate server-side and returns one checked answer instead of raw candidates you have to judge.

resolve_context(ctx, query, memory_type=None, user_id=None)

When to useWhen you want an answer Conxt will stand behind, not a list to sift. It widens the search once if empty, re-ranks by recency if the top hit is flagged conflicting, and says "nothing reliable" rather than guessing. Costs more than get_context — don't make it the default.

Parameter	Type	Default
`query`	string	required	The question to resolve.
`memory_type`	string	null	Restrict to one type (see the type list).
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsOne verified answer, or an honest "nothing found" / "nothing reliable." Bounded to at most one retry.

read

`search_memory`

Targeted semantic search tuned for agent recall. Grep, where get_context is "brief me."

search_memory(ctx, query, memory_type=None, agent_id=None, limit=10, user_id=None)

When to useWhen you know roughly what you're after and want specific hits — optionally narrowed to one type or one agent's writes.

Parameter	Type	Default
`query`	string	required	Search text.
`memory_type`	string	null	Restrict to one type.
`agent_id`	string	null	Only memories written by this agent.
`limit`	int	10 (max 25)	Max results.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsMatching memories, each tagged with its writing agent where known.

read

`search_decisions`

search_memory narrowed to decisions, coding rules, and tool choices only.

search_decisions(ctx, query, user_id=None)

When to useWhen the question is specifically "what did we decide about X" rather than "what do we know about X." Returns the top 10.

Parameter	Type	Default
`query`	string	required	Topic to search decisions for.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsUp to 10 decisions/rules/tool-choices, hybrid-ranked.

read

`recall_as_of`

Recall sorted purely by recency, ignoring reinforcement and confidence entirely.

recall_as_of(ctx, query, as_of_date=None, memory_type=None, limit=5, user_id=None)

When to useWhen a heavily-reinforced old memory might be outranking a newer, correct one — or for point-in-time queries: "what did we believe about this before it changed." Pass as_of_date to look at the graph as of a past date.

Parameter	Type	Default
`query`	string	required	Topic to recall.
`as_of_date`	string (ISO)	null	Only memories on or before this date. Omit for "most recent, period."
`memory_type`	string	null	Restrict to one type.
`limit`	int	5	Max results, newest first.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsTopic matches sorted strictly by creation date, newest first.

read

`get_memories`

Filtered listing with no query — browse and audit, not relevance.

get_memories(ctx, user_id=None, memory_type=None, source=None, date_from=None, date_to=None, limit=20)

When to use"Show me every coding rule," "everything captured from ChatGPT last week." Filter by type, source, and date range.

Parameter	Type	Default
`memory_type`	string	null	One of the seven types.
`source`	string	null	claude · chatgpt · gemini · copilot · grok · perplexity · mcp (aliases accepted).
`date_from`	string	null	Start date — many formats accepted, ISO safest.
`date_to`	string	null	End date.
`limit`	int	20 (max 50)	Max results.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsMatching memories newest-first, or an "invalid source" message listing valid values.

read

`get_pinned`

The always-load set — memories you pinned in the dashboard as permanent context.

get_pinned(ctx, user_id=None)

When to useWhen you want only the handful of rules that should be in context regardless of task. Keep the pinned set short or it stops meaning anything.

Parameter	Type	Default
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsYour pinned memories, or a prompt to pin some at conxt.dev.

Writing

After something is decided — record it so it compounds.

write

`add_memory`

Save a durable fact — a decision, preference, or rule — to your personal graph.

add_memory(ctx, content, memory_type="decision", title=None, tags=None, agent_id=None, session_id=None, user_id=None)

When to useAfter a task produces something worth remembering. One call per distinct fact, so dedup and reinforcement work per record. Routed through type-aware dedup: the same decision saved twice reinforces instead of duplicating.

Parameter	Type	Default
`content`	string	required	The full explanation. Do NOT bury a title, tags, or confidence in here — each has its own parameter, and doing so pollutes the dedup embedding.
`memory_type`	string	"decision"	One of the seven types.
`title`	string	null	Short imperative summary, under 120 chars. Stored as the memory's name.
`tags`	string[]	null	Up to 10 short tags.
`agent_id`	string	null	Your agent name (cursor, codex, claude-code…). Drives the dashboard's attribution badges — pass it consistently.
`session_id`	string	null	Groups memories from one working session.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsConfirmation with scope, type, and a snippet. Note: confidence is not settable over MCP — all writes store at 0.9.

write

`add_team_memory`

Save a memory to a shared team workspace, visible to every member.

add_team_memory(ctx, team_id, content, memory_type="decision", title=None, tags=None, agent_id=None, session_id=None, user_id=None)

When to useWhen a decision generalizes beyond you and your teammates' agents should see it. The bar is higher than add_memory — this is what colleagues rely on.

Parameter	Type	Default
`team_id`	string	required	Target team (get it from list_my_teams).
`content`	string	required	Full explanation — same no-flattening rule as add_memory.
`memory_type`	string	"decision"	One of the seven types.
`title`	string	null	Short imperative summary, under 120 chars.
`tags`	string[]	null	Up to 10 short tags.
`agent_id`	string	null	Your agent name — drives attribution badges.
`session_id`	string	null	Groups a working session.
`user_id`	string	auto	Resolved from the API key.

ReturnsConfirmation showing the verified client and self-reported agent. Requires a real credential — an API key or Supabase session; a bare user_id is refused.

write

`update_memory`

Revise an existing memory in place.

update_memory(ctx, memory_id, content, title=None, tags=None, user_id=None)

When to useTo correct or sharpen a memory. title and tags overwrite only when supplied. Note it repairs the record, not the embedding — materially changed text is better archived and re-added.

Parameter	Type	Default
`memory_id`	string	required	The memory to revise.
`content`	string	required	Replacement text.
`title`	string	null	New title — omit to keep the existing one.
`tags`	string[]	null	New tags — omit to keep existing.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsConfirmation of the update, with an incremented version.

write

`delete_memory`

Retract a memory (soft delete — archived, recoverable).

delete_memory(ctx, memory_id, user_id=None)

When to useWhen something is wrong or shouldn't have been stored. You must own the memory, or be a member of its team.

Parameter	Type	Default
`memory_id`	string	required	The memory to retract.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsConfirmation. The record is archived, not destroyed, and can be recovered.

Teams & agents

Shared workspaces and attribution.

read

`list_my_teams`

List the team workspaces you belong to, with their IDs and your role.

list_my_teams(ctx, user_id=None)

When to useUsually the step before a team read or write — it's where team_id comes from.

Parameter	Type	Default
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsEach team with its name, your role (owner/member), and its team_id.

read

`get_team_memories`

Browse a team's approved shared memories by type. The team analogue of get_memories.

get_team_memories(ctx, team_id, memory_type=None, limit=20, user_id=None)

When to useTo read what a team has agreed on. You must be a member; only approved, non-archived team memories are returned.

Parameter	Type	Default
`team_id`	string	required	The team to read.
`memory_type`	string	null	Restrict to one type.
`limit`	int	20 (max 50)	Max results.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsApproved team memories newest-first, or a not-a-member message.

read

`list_agent_memories`

List memories written by agents — filterable to one agent or one session.

list_agent_memories(ctx, agent_id=None, session_id=None, limit=20, user_id=None)

When to useAttribution and audit: "what has Cursor been saving," or "everything from session run-0705." Only returns memories whose writer is an agent.

Parameter	Type	Default
`agent_id`	string	null	Filter to one agent's writes.
`session_id`	string	null	Filter to one session.
`limit`	int	20 (max 50)	Max results.
`user_id`	string	auto	Resolved from the API key.
`ctx`	Context	injected	Supplied by the runtime.

ReturnsAgent-written memories matching the filter, newest-first.

Parameters shown greyed with a teal default (injected, auto) are supplied by the runtime or resolved from your API key — you never pass them by hand. user_id is resolved from the key on every call; pass it only to act on another identity you're authorized for. Confidence is not settable over MCP; all writes store at 0.9.