Install, use and maintain Elefante. These instructions match the published v2.15.2 release. Your agent does the work; Elefante supplies relevant memory.
FIVE-MINUTE PROOF
Try one real memory.
Use knowledge that matters to your current work. Do not preload a fake conversation or judge Elefante from a button click alone.
Install and restart the agent selected during setup.
Remember one real decision. For example, only when it is actually your rule: “Every customer-facing change must be checked on desktop and mobile before release.”
Ask a later task question: “What must I verify before releasing this customer-facing change?” Have the agent use Recall once.
Inspect the outcome.The decision should be selected and the agent's answer should use it. Selection alone is not enough.
Open Home. Find the same record in Memory Intelligence, run the Recall check, then create a verified backup in Recover.
You need a desktop computer, Python 3.11–3.13, 5 GB free space and internet for setup. The dashboard is included. Git, Node.js and a source checkout are not required.
Use the published installer ZIP, not GitHub’s “Code → Download ZIP”. Extract it before opening the launcher.
macOS: Extract elefante-installer-macOS.zip, then double-click Install Elefante.command. If macOS asks, Control-click it, choose Open, then choose Open again. Administrator access and Terminal commands are not required.
Windows: Extract elefante-installer-Windows.zip, then double-click Install Elefante.bat. The installer guides the rest.
Linux: Extract elefante-installer-Linux.zip, then make install.sh executable and run it from a terminal. The installer guides the rest.
The download page includes the filename and SHA-256 for each package. Compare the archive’s SHA-256 with the published value before opening it.
Step 1
Run the installer
Open the launcher for your platform, shown above, and follow its checks. Read the final result before continuing; starting the installer is not proof that setup completed.
Step 2
Connect your agent and select its workspace
Setup verifies Codex as its required host. Select any additional compatible hosts you use and register your project folder. In v2.15.2, that project separates new memories and Recall—not your dashboard browsing. Elefante connects through MCP; it is not tied to one IDE.
Step 3
Restart your client
Restart the selected host after installation, then ask the agent to check Elefante’s system status.
Step 4
Remember one decision, then check Recall
Ask your agent to remember one real decision for the selected project. Then ask a question that needs it and inspect the memory Recall supplies. Recall returns supporting context, not the final answer; if no memory qualifies, it returns no match.
Choose a specific existing project folder, not your whole home, Documents or Elefante data folder. Register each repository separately if you work across several. A project separates memories; it is not a task manager. Registering a folder does not scan its files or import conversations. Browsing all memories needs no active project; Recall and changes do. This release has no shared-across-project Recall scope.
Remember & Recall
Use your connected agent for everyday work. You do not need to leave the dashboard open.
Remember: ask it to keep one real decision, preference, constraint or lesson. Check whether it saved a record, found an existing one or blocked the write.
Recall: ask a real question that needs that knowledge. Inspect the selected memory and whether it actually helps the answer.
Correct: if the knowledge changes, open its record in Memory Intelligence and review the appropriate change before confirming.
For example, only if this is your rule, ask your agent:
Remember for this project: every customer-facing change must be checked on desktop and mobile before release.
Then ask: “What must I verify before releasing this customer-facing change?” Reuse an existing suitable memory instead of adding a duplicate.
A selected memory or a green status is not proof of a good answer. If a known relevant memory is missed, check its scope and eligibility. Do not store secrets or full transcripts as durable memories. Selected context sent to your agent is subject to that provider’s policies.
Dashboard guide
On the computer where Elefante is installed, open localhost:8000. No browser extension or connector is required. If it does not open, ask your connected agent to Open Elefante Home and use the address it returns.
This is a maintenance console, not a chat window. Most browsing reads a snapshot of stored data. An operation receipt is a result card showing what happened and what was verified. A managed change, backup or restore should finish with VERIFIED_COMPLETE.
Reload snapshot rereads the existing snapshot; it does not rebuild it from the stores. Current describes its age, not memory accuracy. Light / Dark changes appearance only. If controls expire, use Reconnect Home, then explicitly retry or review a fresh change plan.
Home — status and the next useful check
See memory counts, review signals, the active boundary and recovery evidence. Continue opens the recommended action: Review opens its queue; Test Recall opens Recall.
“Review 10 direct signals” means ten records have health or lifecycle information worth inspecting. It does not mean ten records are false. Read each reason and leave correct knowledge unchanged. Empty Home recommends remembering one useful decision.
Recall — inspect what memory your agent receives
Check the displayed scope, enter a real question and select Run Recall Check. Open a selected record to read it. This needs connected controls and an active project whose folder is available.
Bundle supplied: inspect whether the selected memories help.
No memories selected: nothing qualified. Use Inspect available memories if you expected a match.
Blocked: read the conflict or governance reason before changing knowledge.
Unavailable: the check did not complete. Check the connection; this is not a successful no-match.
The question is not saved as a memory. The question and result survive tab switches, but a page reload or project change clears them. Missing counts are not zero; a missing valid verification time means Not verified.
Memory Intelligence — inspect and correct records
Library browses the snapshot across projects. Review shows direct health or lifecycle signals. Select a row to read content, source, scope, health and relationships. Close panel returns to the list on narrow screens.
Snapshot search: text matching after two characters, not governed Recall. Zero matches means zero rows. Filter memories narrows the displayed rows.
Edit: fix wording without changing meaning. Replace: keep newer knowledge while preserving the old version.
Archive: stop the record being eligible for Recall without deleting it. Restore: return a manually archived record that has not been superseded.
Verified Resolve: for a represented conflict pair, choose the authoritative record. Elefante does not decide truth for you.
Delete permanently: erase after explicit confirmation and verification. Older backups may still contain the memory.
Available changes depend on the record and active project. Review the plan, confirm deliberately and inspect the final receipt. At risk is a reason to inspect, not an instruction to delete.
Connections — understand topics and relationships
Topics groups represented memories by subject. Vitality shows stored maintenance scores, not truth or usefulness. Decision Graph follows explicit stored relationships.
Similar wording does not create a connection. An empty graph can be correct when no qualifying relationships are represented. This view is read-only; it does not prove that memory improved a task.
Projects — choose the memory boundary
Register specific folders and review older unassigned memories before assigning them. Do not guess ownership. After registration and strict-mode setup, choose Use for actions beside the intended project; Current scope confirms it.
Changing project clears the old Recall question and receipt. It does not move memories, import files or repeat an operation. Disabled management controls explain their missing requirement.
Recover — diagnose, back up and restore
Check health diagnoses without repairing. Back up now creates and verifies a local backup. Restore first inspects a chosen backup and asks you to confirm a plan; it replaces stored state. Do not casually test it on real memories.
Support report previews what will be included, then creates a local privacy-safe ZIP. It does not send the ZIP anywhere. Read each final receipt; a failed or unsafe result is not completion.
Repair, update, product rollback and uninstall use the matching official package, not dashboard action buttons.
Usage & privacy
Open Home → Advanced → Session Intelligence. This reports usage; it does not start collection. After explicit permission, Elefante automatically records local MCP activity and token estimates, without prompts, transcripts or response content.
Recorded events: activity, including failed and blocked calls—not completed tasks. Refreshing the report adds no event.
Usage cost: needs complete provider-reported usage and matching rates. Estimates stay separate. Unavailable does not mean free.
Task result: needs outcome evidence. Not verified means no useful result has been demonstrated by that evidence.
Usage details: expands counts, coverage, timestamps and limits. Suggestions are provisional, not proven improvements.
These activity records are separate from memories in the Library. Permission, retention, export and deletion use the installed package’s command-line controls, not dashboard or MCP buttons. Read the privacy notice.
Local usage controls — advanced
Run these only for the intended installation. On macOS or Linux, open a terminal in ~/.elefante/app/current. On Windows, open PowerShell in %LOCALAPPDATA%\Elefante\app\current and replace ./.venv/bin/python below with .\.venv\Scripts\python.exe.
Check status without changing collection:
./.venv/bin/python scripts/pipeline/session_intelligence.py status
Use the same command with snapshot to refresh the report or revoke to stop new writes. Revoking does not delete retained evidence. Provider usage and aggregate hypotheses require their own permission.
Default event retention is 30 days, applied when the ledger opens or is explicitly pruned. For export, deletion and pruning, inspect the installed command’s --help first; deletion requires an exact target and confirmation. These controls do not change semantic memories or graph data.
USER CAPABILITY MAP · v2.15.2
What the published product can do.
These capabilities belong to the downloadable v2.15.2 release. Some require explicit permission or operator action; a capability being available does not mean it runs automatically. Unknown usage or cost remains unknown.
CORE RELEASE
Bounded Recall
One read-only call returns a small governed context bundle—or explicitly abstains when evidence is irrelevant or unsafe.
Use
Ask one real task question through your connected agent, or run the dashboard Recall check.
Behavior
A small eligible memory bundle, or an explicit no-match, blocked, or unavailable result.
CORE RELEASE
Strict project isolation
New memory and Recall stay inside one active registered project; missing or invalid isolation evidence fails closed instead of sharing across projects.
Use
Choose one specific registered folder for memory actions.
Behavior
New memory and Recall stay inside that boundary; the folder is not scanned or imported.
CORE RELEASE
Verified correction
Edit, Replace, Archive, Restore, Resolve, and backup-bound permanent deletion verify storage, Home, and scoped Recall, with compensation when a postcondition fails.
Use
Open a record in Memory Intelligence and choose Edit, Replace, Archive, Restore, Resolve, or permanent deletion.
Behavior
A reviewable plan, deliberate confirmation, and a verified completion or failure receipt.
CORE RELEASE
Verified recovery
Health, verified backup, rollback-protected restore, and a privacy-safe support ZIP return bounded receipts without exposing memory content.
Use
Open Recover to check health, back up, preview and confirm restore, or create a support report.
Behavior
A bounded local operation with postcondition evidence and rollback protection where required.
OPT-IN OR EXPLICIT
Conflict-safe memory
Detected conflicts can be withheld from Recall; deterministic detection and dry-run Smart Merge keep resolution under explicit authority. Detection is not a guarantee that every contradiction is found.
Use
Inspect the represented records before resolving a conflict or applying Smart Merge.
Behavior
Unsafe conflicting records can be withheld; Elefante does not decide truth for you.
OPT-IN OR EXPLICIT
Retrieval Explanation
Snapshot search exposes available match, ordering, source and health evidence. It is not a complete explanation of governed MCP Recall or proof of answer quality.
Use
Use snapshot search, Review, and a record's source and health details.
Behavior
Inspectable snapshot evidence—not a complete explanation of governed Recall or proof of answer quality.
OPT-IN OR EXPLICIT
Session Intelligence
With explicit permission, automatic MCP capture records local activity and token estimates. Supplied provider-reported usage stays separate from estimates. No prompts, transcripts, or response content are stored; unsupported cost and task value remain unknown.
Use
Grant a specific local permission, then inspect Home → Advanced → Session Intelligence.
Behavior
Metadata-only activity and usage evidence; unsupported cost and task value remain unknown.
OPT-IN OR EXPLICIT
Live Distiller
An opt-in foreground watch mode processes supported session files serially and stores nothing unless storage is explicitly enabled.
Use
Run the Session Distiller in the foreground and opt into watch or storage explicitly.
Behavior
Supported session files are processed serially; nothing is stored unless storage is enabled.
OPT-IN OR EXPLICIT
Team Sync
Signed, scope-bound local bundles exchange an explicit memory allowlist without introducing cloud transport.
Use
Use the advanced local CLI with an exact memory allowlist and a user-chosen transport.
Behavior
A signed, scope-bound export or additive import; no cloud transport is provided.
OPT-IN OR EXPLICIT
Local media attachments
Bounded image, audio, and video attachments remain content-addressed and local, with no OCR, transcription, model analysis, or network upload.
Use
Attach a bounded local image, audio, or video file through a memory operation.
Behavior
Local content-addressed storage and integrity evidence—no OCR, transcription, analysis, or upload.
OPT-IN OR EXPLICIT
Private host events
Typed, privacy-scrubbed file, terminal-error, and conversation events can request governed retrieval without persisting host content.
Use
Have an integrated host send a typed, privacy-scrubbed event to the local endpoint.
Behavior
Literal-trigger retrieval without silent interception or persistence of the event body.
OPT-IN OR EXPLICIT
Elefante Home
An advanced loopback-only maintenance console inspects the redacted snapshot and exposes only named, short-lived, verified controls; it is not the daily agent workflow.
Use
Open localhost:8000 on the installed computer or ask the connected agent to open Elefante Home.
Behavior
A local maintenance and inspection console; it is not the everyday chat surface.
Compatible agents
Codex is the primary validation target. Other listed hosts have released compatibility adapters, not vendor certification. IBM Bob and Antigravity remain Preview; Agent Zero has a documented community connection path.
VS Code — Released adapter · optional compatibility, not vendor-certified
Cursor — Released adapter · optional compatibility, not vendor-certified
Kiro — Released adapter · optional compatibility, not vendor-certified
Gemini CLI — Released adapter · optional compatibility, not vendor-certified
Claude Code — Released adapter · optional compatibility, not vendor-certified
Codex — Primary validation target · not vendor-certified
OpenClaw — Released adapter · optional compatibility, not vendor-certified
Zed — Released adapter · optional compatibility, not vendor-certified
Continue — Released adapter · optional compatibility, not vendor-certified
Troubleshooting
Home will not open
Confirm Elefante is installed on this computer. Ask your connected agent to open Home and check system status.
Setup required or controls unavailable
Read the stated reason. Reconnect an expired session; use Projects if the boundary is missing. A read-only snapshot alone is not a reason to reinstall.
Recall misses a known memory
Check Library, active project, archive or supersession state, conflicts and the question. A snapshot text match does not guarantee Recall eligibility.
At risk, but the content is correct
Read the health and lifecycle reason. No correction may be needed.
The graph is empty
Check for explicit relationships. Topic similarity and a nonzero memory count are not enough.
The snapshot is still old after Reload
Reload rereads the same file. Ask your connected agent to refresh or open Home. Do not delete storage or remove lock files.
Usage cost or task result is unavailable
Open Usage details for missing evidence. Do not substitute estimates for measured results.
For persistent problems, use Recover → Support report. Review the contents before sharing. Do not reset memories, edit databases or dump credentials and logs as a first response.
Repair & uninstall
Rerunning the same official installer repairs its runtime while preserving durable data. Check the version in your dashboard first; a complete uninstall requires the official package that exactly matches the installed build.
macOS: open Uninstall Elefante.command.
Windows: open Uninstall Elefante.bat.
Linux: from the extracted matching package, run chmod +x uninstall.sh && ./uninstall.sh.
The package shows the impact and asks you to type UNINSTALL. It verifies a restorable backup, removes the app and unchanged Elefante-owned connections, and checks that memories were preserved. Your durable data is not deleted.
Inspect the completion receipt. If the package reports a mismatch or unsafe state, stop and use the support report. A later matching installation can reconnect the preserved data; treat that reconnection as complete only after Doctor and a live Recall check pass.
Advanced reference
Most customers can stop after Remember, Recall, correction, and backup. The released interface also exposes structured surfaces for agents and operators. They stay separate from the first-use path.
Context, sessions, tasks, ETL, and directives
These surfaces are for agents or operators that need structured control beyond the normal Remember and Recall loop.
Context and Sessions: retrieve a broader graph-connected bundle or inspect time-based work sessions.
Tasks: preserve task hierarchy, dependencies, status, and output. This is separate from project memory isolation.
ETL: enrich stored memories with summaries, concepts, and explicit trigger metadata.
Directives: manage always-on behavioral constraints that must not depend on similarity ranking.
Prompts: the two published prompts provide grounding or pre-fetched memory context; they are not tools.
Where exact schemas and source live
This page contains the complete customer path and every released feature group. Exact tool parameters, response schemas, scoring formulas, architecture, source, and development history live in the GitHub repository linked once in the site footer.
LOCAL ARCHITECTURE
The daemon owns the data boundary.
A loopback-only daemon owns storage, concurrency, and provenance. MCP clients connect through local HTTP or a storage-free stdio bridge; they do not become competing database owners.